RISC-V Daily Intelligence Report (2026-11-16)
Research window: 2026-11-13 05:00 ~ 2026-11-16 05:00 (Beijing time, 72 hours) Sources: Google News aggregation (24 query sets in Chinese and English, via proxy), Hacker News, Phoronix, CNX Software, LinuxGizmos, RISC-V International official site, openRuyi official site and main repo, RISC-V member company website sweep (391 companies), riscv100.top leaderboard sweep, plus commit and release searches on GitHub across organizations and key repositories including XiangShan/NEMU/difftest/XSCache, CORE-V CVW/CVA6, DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT and spacemit-robotics, Sipeed, revyos, openRuyi-Project, and Radxa Linkr.
This Issue’s Index
- Today’s Highlights: Purism unveils Librem 5 v2, octa-core SpacemiT K1 at a $1,599 crowdfunding price (11-14)
- I. Open-Source Community and Software Ecosystem
- Full picture of RISC-V matrix extension roadmaps: IME, VME, AME and dot-product extensions side by side (11-15)
- FURSOY open-source out-of-order RV64 core: boots Linux 6.6.50, closes timing at 500MHz on 45nm (11-15)
- Twenty openRuyi commits: LLVM enables PIE across three lines, python-torch-ruyiai lands in tree (11-14)
- ReVyOS unifies kernel config: six platform configs and linuxboot enabled in one pass (11-15)
- Three XiangShan ecosystem repos: NEMU fixes MMU and CBO, XSCache fixes L2 counting (11-15)
- IAR summit recap: one toolchain for fragmentation, functional safety covering ten standards (11-13)
- AI-assisted design compresses RISC-V IP customization cycles: Andes case drops from 8.5 to 6 weeks (11-13)
- Desktop systems on ESP32-P4: the Sylphide OS and BMDS routes (11-14)
- II. Product and Ecosystem Developments
- Longqing Aerospace LQ7000: domestic RISC-V paired with in-house GPNPU, 20 to 400 TOPS (11-13)
- FIORA5 open equity crowdfunding: a Romanian RISC-V edge AI processor (11-14)
- India’s VVCE 65nm RISC-V teaching chip sent to TSMC for tapeout (11-15)
- III. Deep Dive on Domestic Teams
- Lingrui Zhixin Pre-A+ round in the hundreds of millions of RMB: P100’s SMT4 and 128-core timeline (11-13)
- China Business Journal: RISC-V main control chip market grows from RMB 1.1 billion to RMB 37.7 billion in five years (11-14)
- Sophgo sophon-tools eleven commits: management-plane fixes and default credential remediation (11-15)
- IV. Special Chapter on Embodied Intelligence and Robotics
- Beijing’s three-year embodied intelligence action plan puts RISC-V intelligent computing chips on its breakthrough list (11-11)
- Compute tiering for on-device agents: LQ7000’s TOPS and P100’s SMT4 each cover a layer (11-15)
- V. Summary and Trend Observations
- Appendix: Complete Source List and Source Verification Table
Today’s Focus: Purism Unveils Librem 5 v2 with Eight-Core SpacemiT K1 and $1,599 Crowdfunding Price (11-14)
Purism launched crowdfunding for the Librem 5 v2 on November 14, marking the first time the Librem 5 series has switched its main processor from Arm to RISC-V. The official page lists the following specs: eight-core SpacemiT K1, 5.7-inch 720×1440 display, 8GB RAM, 128GB storage, 4500mAh replaceable battery, dual SIM, 8MP front and 13MP rear cameras, three hardware kill switches for cellular, Wi-Fi/Bluetooth, and microphone/camera, and PureOS based on Debian, capable of running Android apps via the Waydroid compatibility layer. The crowdfunding price is $1,599, with a goal of securing 2,500 orders by February 15, 2027; test units are expected in late 2027, with mass production delivery targeted around May 2028. LinuxGizmos, OMG! Ubuntu, and Liliputing followed up between November 14 and 16, with reporting consistent with the official page.
The significance of this news lies in “who is switching architectures.” RISC-V phones have mostly been geek projects or small-batch products in development board form, whereas Purism is one of the few Linux phone vendors with a continuous product line, its own distribution, and an existing user base. By betting its next-generation flagship directly on the K1, it is effectively acknowledging that RISC-V application processors can now support the positioning of “a Linux phone for daily use,” rather than just serving as validation boards. For SpacemiT, the K1 entering both Start9’s router and Purism’s phone within two weeks shows that this chip is penetrating from the development board market into finished-product brands—a rare case of two parallel finished-product lines for a domestic RISC-V SoC.
The official page also lays out the risks plainly: RISC-V is not yet a “plug-and-boot” platform for PureOS, the graphics stack can currently only display images, and full 3D support and optimization still require rewriting and investment—which is precisely why crowdfunding is needed to prove demand before organizing development. In other words, the bottleneck for this phone is not the K1’s compute power, but the mobile graphics stack, camera pipeline, and power management—the parts that are “ready-made in the Arm ecosystem but must be rebuilt on RISC-V.” The 2028 delivery timeline, the 2,500-order crowdfunding threshold, and the delivery risk inherent in the crowdfunding model itself are all hard constraints for observing whether this machine can ship.
Placing it alongside other news from this week makes the pattern clearer: around the same day, FIORA5 discussed a 2027 prototype and 2029 licensing, while Lingrui Zhixin’s 128-core server CPU targets 2027 tape-out and 2028 mass production. High-value finished products and high-performance chips in RISC-V generally have timelines falling between 2027 and 2029; the two years in between are a window for software stacks and post-silicon validation, not a window from launch event to store shelf.
Date: 2026-11-14 Sources: Purism: Librem 5 v2 crowdfunding page, LinuxGizmos: Librem 5 v2 adopts eight-core RISC-V, OMG! Ubuntu: Second-generation Librem 5
I. Open Source Community and Software Ecosystem
RISC-V Matrix Extension Landscape: IME, VME, AME, and Dot-Product Extensions Side by Side (11-15)
FPRox published a survey of matrix extension status on November 15, comparing several proposals under discussion at RISC-V International on a single table organized by three multiply-accumulate layouts: “dot product, fat inner product, outer product.” The proposals listed include: the four-element dot-product extension Zvdot4a8i for 8-bit integers, the Integrated Matrix Extension IME (Zvv*mm) advanced in a separate repository, VME (Zvt) attached under the vector extension, the standalone Attached Matrix Extension AME (Ztt), and batched dot product Zv[b]dota. The author states the specification status is as of mid-November 2026 and will be updated as drafts change.
The difference among the three layouts is not “whether it can compute” but arithmetic intensity and register pressure. Dot product is the most flexible—vector-by-matrix and matrix-by-matrix can both be decomposed into dot products—but each computation requires moving a linear number of elements, giving the lowest arithmetic intensity. Fat inner product is closest to classic matrix multiplication and suits matrix-by-matrix, but has low utilization for vector-by-matrix. Outer product compresses the inner-product dimension to near 1; as long as the accumulator stays resident and the k dimension is large enough, linear loads buy quadratic computation, giving the highest arithmetic intensity—but with the strictest operand-shape requirements. These differences determine whether to add a new register file or reuse vector registers, how to arrange memory layout, and whether compilers can automatically map existing operators onto them.
For the industry, the real risk is not the number of proposals but their incompatibility. IME uses a dedicated matrix register file and dedicated instructions, VME depends on vector registers, and AME is yet another attachment scheme; if these do not converge on a single mainline, operator libraries and compiler backends will have to maintain a path for every combination, weakening RISC-V’s selling point of “one codebase running across multiple chips” for AI inference. Conversely, this also explains why, in the same week, IAR discussed toolchain fragmentation, XiangShan discussed verification correctness, and distributions discussed platform unification—all different facets of the same problem.
Date: 2026-11-15 Source: FPRox: RISC-V Matrix Extensions
FURSOY Open-Sources Out-of-Order RV64 Core: Boots Linux 6.6.50, Closes 500MHz on 45nm (11-15)
FURSOY published fursoy-cpu on November 15, codenamed Fursoy Slice-1: a 64-bit out-of-order superscalar RISC-V processor written from scratch in SystemVerilog, with an ISA of RV64IMAFDV plus Zba, Zbb, Zicsr, and Zifencei. The measured claims are that it can boot mainline Linux 6.6.50 and BusyBox user space under Sv39 virtual memory, and that under Nangate 45nm standard-cell synthesis, all eight core functional blocks close timing at 500MHz. The repository uses the Apache 2.0 license and, besides RTL, includes SBI firmware, device trees, Linux 6.6.50 configuration materials, a Verilator testbench, and Python scripts for lint, regression, Sail differential co-simulation, synthesis checks, and the Linux bring-up flow.
The core is not positioned against server-class IP but rather aims to make the full chain of “out-of-order + virtual memory + runs Linux” a reproducible open-source project. Many open-source cores can boot Linux, but few simultaneously provide synthesis timing closure results, Sail differential verification, and complete build scripts; for teaching and research teams, this means they can modify the microarchitecture from a working baseline without first spending months setting up an environment. The 2-wide issue, 45nm process, and 500MHz frequency also indicate it is still some distance from production IP on PPA—its value lies mainly in methodology and the verification loop.
Viewed alongside XiangShan, open-source cores at home and abroad are stratifying: XiangShan’s Kunminghu takes the high-performance and industrial-grade verification route, while projects like FURSOY take the “small but complete, reproducible” route. Both routes need the same infrastructure, including differential testing, Sail models, and repeatable synthesis scripts—which is why this week’s difftest and NEMU commits are worth recording.
Date: 2026-11-15 Source: FURSOY/fursoy-cpu, LavX News: FURSOY Open-Source Core Boots Linux and Closes 500MHz in Synthesis
openRuyi’s Twenty Commits: Three LLVM Lines Enable PIE, python-torch-ruyiai Enters the Tree (11-14)
The openRuyi main repository landed twenty commits in the window, concentrated from the evening of November 13 to the afternoon of November 14. Three LLVM lines (22, 23, 24) enabled PIE in the same batch and generated clang resource directories, clang-wrap tracked the 20261113 snapshot, and cmake was bumped to 4.4.5; the python-torch-ruyiai package was added; on the KDE side, bluedevil and kf6-bluez-qt were added; and a batch of Go ecosystem packages entered the tree, including go-maunium, go-go-mau-util, the go-modernc series, and pion’s turn and transport—judging from dependencies, this is paving the way for a Matrix client stack. At 15:13 on November 14, one commit each for linux and linux-lts landed in sync with “refresh patch set + refresh config.”
Enabling PIE across all three LLVM lines is a distribution-level hardening move: PIE loads executables at randomized addresses and is a prerequisite for ASLR to take effect; enabling it line by line shows all three versions will be maintained long-term rather than only one mainline. The entry of python-torch-ruyiai into the tree is more noteworthy—it is the first time the RuyiAI software stack appears as a distribution package, meaning RuyiAI’s PyTorch adaptation no longer exists only in a standalone repository but is entering an installable, dependable distribution path, a necessary step for downstream users to treat RVA23 machines as AI development machines.
The Matrix-related Go packages and KDE Bluetooth components both point to desktop usability: the former fills out the dependency tree for a communications client, the latter fills in Bluetooth management for the desktop environment. Together with the previous day’s kernel patch set refresh, openRuyi had no release in this window—what it did was “push the AI stack, desktop stack, and kernel baseline forward simultaneously,” consistent with its rolling-release positioning.
Date: 2026-11-14 Source: openRuyi-Project/openRuyi
ReVyOS Unifies Kernel Config: Six Platform Configs and linuxboot Enabled at Once (11-15)
ReVyOS’s linux repository landed twenty commits in a burst on the evening of November 15, all submitted by the same maintainer. On the config side, debian_defconfig was added, enabling common options and the Picoheart errata together, and separately enabling six platform configs: Sophgo SG2042 and SG2044, T-Head TH1520, SpacemiT K1 and K3, UltraRISC DP1000, and ESWIN EIC7700, plus a linuxboot config and cross/native CI builds. On the fix side: an errata replacing WRS.NTO with WRS.STO on K3, MMC card clock and GPIO card-detection fixes for EIC7700, and synchronization of the initial DMA mapping for DRM xe on non-coherent devices.
Covering six platforms with one config represents a shift in distribution maintenance. Previously each platform had its own defconfig, so every kernel upgrade required reworking each one, making it easy to miss a board or apply an errata to only one config; after unifying to one baseline plus platform switches, platform differences collapse into config options, and CI can verify both cross and native lines simultaneously. For downstream vendors, this means “whether my board is taken care of in the kernel config” can be checked at a glance rather than digging through some branch’s history.
Two hardware fixes deserve separate mention: the difference between WRS.STO and WRS.NTO on K3 concerns storage timeout semantics—a typical post-silicon errata that only surfaces under real workloads; the EIC7700 MMC clock requires stopping the card clock before changing the rate and clearing CDTEST on reset—issues that directly affect whether the SD card can be reliably detected. Neither is a new feature, yet both determine whether a development board can serve as a daily machine.
Date: 2026-11-15 Source: revyos/linux
XiangShan Ecosystem’s Three Repos: NEMU Fixes MMU and CBO, XSCache Fixes L2 Counting (11-15)
Commits across the XiangShan ecosystem in this window all landed on the correctness and verification side. NEMU had four: scoping final address translation cache invalidation by stage, synchronizing dirty FS and VS state to Spike, invalidating LR/SC reservations when cache management instructions (CMO) execute, and enabling the Zfh and Zfhmin half-precision floating-point extensions in the kmhv2 config. The XiangShan main repo had two: splitting performance CI build and run into two steps, and relaxing the AI usage reminder in the docs while clarifying commit and PR guidelines. XSCache had one: fixing a carry-loss bug in CoupledL2’s TX reservation counter. difftest had one: changing the fork checkpoint to 1 second and the packet slot to 2 to the 22nd power.
These four NEMU changes point to the same thing: boundary conditions in virtualization and cache coherence. If address translation cache invalidation is not scoped by stage, VS-stage invalidation will wrongly hit FS-stage entries, and guest and host mappings will contaminate each other; if LR/SC reservations are not invalidated after CMO, atomic operations may commit successfully based on a cache line that has already been cleaned—an error very hard to hit in random testing; synchronizing dirty state to Spike lets the differential test’s reference model and the model under test compare on the same baseline, otherwise differences get misjudged as RTL bugs.
The two commits in the main repo and cache repo are more like engineering governance: splitting CI steps makes performance regression failures easier to localize, and the L2 carry loss is a counter bit-width detail that only manifests as statistical bias under long stress tests. Overall, the focus at this stage of Kunminghu V3 is nailing down the correctness of virtualization, atomic operations, and cache management rather than adding new features.
Date: 2026-11-15 Source: OpenXiangShan/NEMU, OpenXiangShan/XiangShan, OpenXiangShan/XSCache
IAR Summit Recap: One Toolchain Against Fragmentation, Functional Safety Covering Ten Standards (11-13)
EE Times China published IAR’s recap of the 6th RISC-V China Summit on November 13. The article groups embedded development pain points into three: architecture fragmentation leading to toolchain fragmentation, diverse functional safety standards, and long product lifecycles. The corresponding solution is a unified IDE plus underlying compiler, debugger, and linker scripts covering both Arm and RISC-V, so teams keep the same debug platform, compiler options, and script formats when switching architectures; IAR says it launched a commercial RISC-V toolchain as early as 2019, and supports the not-yet-finalized P extension by vendor version to accommodate different encoding formats.
The functional safety section is where this article carries the most information. It explains the relationships among standards as a chain of references: the industrial robot system standard ISO 10218 references the methodology standard ISO 13849, which at the software tool level references the foundational standard IEC 61508; clarifying this chain gives toolchain safety arguments a clear basis. IAR says its functional safety edition covers ten standards across automotive, industrial, medical, and rail transit, and provides safety guides and certification documents. On the debug side, it mentions multicore and heterogeneous multicore debugging, Trace, and letting AI agents call the debug platform via the Model Context Protocol.
This article is essentially a vendor perspective, but it states an industry fact: RISC-V fragmentation ultimately gets backstopped by toolchain vendors. Chip vendors can customize instructions and operating systems can do their own adaptation, but an industrial controller meant to sell for ten years must have a certified compilation and debug environment with long-term support—which is also why open-source toolchains are hard-pressed to replace commercial ones in functional safety scenarios in the near term.
Date: 2026-11-13 Source: EE Times China: RISC-V China Summit Recap, One Toolchain Against Architecture Fragmentation and Functional Safety Challenges
AI-Assisted Design Compresses RISC-V IP Customization Cycle: Andes Case from 8.5 Weeks to 6 Weeks (11-13)
EE Times published a report on Architect Labs on November 13: the company used an AI system to complete the full flow of a low-power, low-latency physical AI accelerator from spec to verification and deployment in under two weeks, with human architects writing the spec and AI generating RTL, UVM verification, formal proofs, firmware, drivers, and kernel; the first test chip (A0) succeeded on the first try. The report also provides competitive context: ChipAgents claimed in September that it cut an Andes processor customization task from 8.5 weeks to 6 weeks, Ricursive’s founders come from Google Brain’s AlphaChip team, and Synopsys is also using AI agents to handle design complexity.
RISC-V’s place in this narrative is specific: customization is its core selling point, and the high cost of customization is its core obstacle. Customers wanting to add custom instructions, modify pipelines, and pair dedicated accelerators used to need a team for months; if AI can compress IP customization to weeks, more vendors will be willing to modify cores for specific workloads, and RISC-V’s “extensibility” will truly become a commercial advantage. The significance of the Andes case is not the absolute difference between 8.5 and 6 weeks but that it shows the customization step is beginning to be tooled.
The skepticism to retain: these numbers all come from vendor self-reports, lacking independent benchmarks and post-tapeout production data; the two-chip comparison is also calculated under a specific process projection. For observers, what can be confirmed at this stage is that the flow is being compressed, not that the compressed designs hold up equally on real workloads and yield.
Date: 2026-11-13 Source: EE Times: Inside Architect Labs’ Two-Week Chip Design
Desktop Systems on ESP32-P4: Two Routes, Sylphide OS and BMDS (11-14)
CNX Software reported on November 14 on two projects bringing desktop systems to microcontrollers, both targeting ESP32-P4 boards with HDMI, Ethernet, and USB. Sylphide OS, developed by Thomas Capricelli, runs on FreeRTOS underneath with an interface mimicking the early Macintosh, already featuring a desktop, menus, window management, icons, and an info panel, with drivers covering mouse, keyboard, display, storage, and network, under GPL v3; besides ESP32-P4 it also supports Raspberry Pi Pico 2 and FabGL boards, and provides an emulator on Linux. The other project, Bare Metal, Do Stuff, takes a DOS-style route, also targeting ESP32-P4.
ESP32-P4 is Espressif’s dual-core RISC-V high-performance MCU with PSRAM and display interfaces, and has been the main vehicle for “using a microcontroller as a computer” experiments these past two years. Sylphide OS’s approach keeps the real-time kernel at the bottom and builds graphics and window management on top, effectively reimplementing a desktop paradigm on an RTOS; BMDS returns to a simpler single-task or lightweight multitasking model. What the two routes share is that they do not pursue general-purpose computing power but spend limited RAM and display bandwidth on interface responsiveness.
For the RISC-V ecosystem, the value of such projects lies on the software side: they force developers to re-polish foundational pieces—framebuffers, bitmap fonts, window clipping, and input events—on resource-constrained RISC-V platforms. This code will not go directly into mass production, but it will become reference implementations for future RISC-V MCU graphics solutions.
Date: 2026-11-14 Source: CNX Software: Desktop Operating Systems for ESP32-P4-Based PCs
II. Product and Ecosystem Updates
Dragon Engine Aerospace LQ7000: Domestic RISC-V Paired with In-House GPNPU, 20 to 400 TOPS (11-13)
Dragon Engine Aerospace released a white paper on its agent processor technology framework on November 13, alongside its first edge AI processor, the LQ7000. According to the official statement: the chip uses a domestic RISC-V plus in-house EE-DNA GPNPU architecture, with peak compute spanning 20 TOPS to 400 TOPS, supporting INT4, INT8, and FP16, and preloaded with mainstream large models such as Qwen, DeepSeek, and LLaMA, targeting home, education, office, and government scenarios; there is also an in-orbit AI line, claimed to push remote sensing from “space sensing, ground computing” to “space sensing, space computing,” with target information utilization for in-orbit large-model inference exceeding 90%. On the same day, the company entered an ecosystem strategic partnership with ChipON and signed an edge AI product strategic cooperation and procurement intent agreement with Liulian Intelligence.
What is worth noting is the product definition approach: a single chip quoting a range of 20 to 400 TOPS indicates it is selling a configurable IP combination rather than a single SKU, with customers selecting tiers by power and cost. RISC-V here handles control and scheduling, while matrix computation is delegated to the in-house GPNPU—a “general-purpose core + dedicated array” division of labor consistent with most domestic edge AI chips, the difference being whether the general-purpose core is replaced with RISC-V.
What warrants caution is the level of disclosure: the launch event gave no mass-production or shipment timeline, the TOPS range was not mapped to specific models and power figures, and the 90% utilization claim for in-orbit scenarios lacks third-party verification. ChipON’s participation and Liulian Intelligence’s procurement intent are the currently verifiable progress; the actual mass-production milestone remains undisclosed.
Date: 2026-11-13 Source: MyDrivers: Dragon Engine Aerospace Releases Edge AI Compute Chip LQ7000, Eastmoney: Dragon Engine Aerospace Releases LQ7000
FIORA5 Opens Equity Crowdfunding: Romania’s RISC-V Edge AI Processor (11-14)
Romanian deep-tech startup FIORA5 opened its first equity crowdfunding round on the Growceanu platform, with participation starting at EUR 200, after Vest Ventures and private investors had already invested. Based in Oradea, the company develops configurable RISC-V processor IP for edge AI, focused on low latency, deterministic real-time execution, and low power, targeting devices including smart glasses and edge inference terminals. According to the company, its core reduces on-chip area by up to 50% and energy consumption by up to 60%; it plans to produce a functional prototype in 2027 and an licensable architecture by 2029. Vest Ventures completed its first direct investment of EUR 150,000 in the company in October.
This is a typical sample of European RISC-V startups: no complete devices, no SoCs—only licensable processor IP, pooling small amounts through crowdfunding and filling gaps with industry funds. The advantage of this model is high capital efficiency and a clear narrative; the downside is an extremely long validation cycle—from IP to customer tape-out to mass production, any delay in any link pushes the 2029 licensing target further out.
The area and energy figures should be discounted: they are self-reported values against “existing solutions,” without specifying the comparison target, process node, or workload. For early-stage companies like this, what is more worth tracking is whether it can find the first customer willing to put its IP into a mass-production chip, rather than the percentages at a launch event.
Date: 2026-11-14 Source: Foro3D: FIORA5 Opens Funding Round on Growceanu for Its RISC-V Edge AI Processor, ITKeyMedia: Vest Ventures Invests in FIORA5
India’s VVCE 65nm RISC-V Teaching Chip Sent to TSMC for Tape-Out (11-15)
The Hindu reported on November 15 that VVCE engineering college in Karnataka completed the design of an in-house RISC-V-based chip using a 65-nanometer process, sent to TSMC for tape-out on November 13. The project was led by the college’s Department of Electronics and Communication Engineering in collaboration with AstraSemi and KYROS SEMI; the institution says its significance lies in exposing students and researchers to the complete advanced semiconductor design and tape-out flow. The report did not disclose the chip’s specific specifications, core count, or purpose.
The combination of 65nm and a teaching project indicates this is not a performance contest but training in talent and process. India’s Chips to Startup program and design incentive schemes have connected hundreds of institutions and dozens of startups to shared EDA and multi-project wafer services; college-level tape-outs are a product of this infrastructure. For RISC-V, students who complete the full design-to-tape-out flow using an open-source instruction set while in school will face a much lower barrier to entering the industry after graduation.
Such news looks minor in isolation, but it is an indicator of ecosystem depth: a country’s RISC-V talent pool is not built by a few corporate launch events, but by cohort after cohort of students actually sending chips to tape-out. What can be tracked going forward is the post-silicon test results of this chip and whether it yields reusable teaching IP.
Date: 2026-11-15 Source: The Hindu: VVCE Achieves Milestone in Indigenous Chip Design
III. In-depth Look at Domestic Teams
Lingrui Zhixin Pre-A+ Round of Several Hundred Million Yuan: P100’s SMT4 and 128-Core Timeline (11-13)
Shanghai Lingrui Zhixin announced on November 13 that it had completed a Pre-A+ round of several hundred million yuan, led by Meituan’s strategic investment arm, with participation from a JD-affiliated fund, Ganbei Fund, Lihe Suanxin Fund, Junxi Capital, SenseTime Guoxiang Capital, Shangqi Capital, and Amber Capital, while existing shareholders continued to increase their stakes. The company was founded in early 2025, is registered in Zhangjiang, Shanghai, and has raised close to one billion yuan cumulatively. On the product side, the core is the server-class RISC-V CPU core P100 released in January this year: a single core supports dynamic simultaneous multithreading SMT4, switchable between SMT1 and SMT4 without rebooting the system, paired with a 20+ stage out-of-order pipeline and 8-decode 10-issue; in August this year a test chip based on P100 was taped out on a domestic advanced process line, and in September it was disclosed to have entered the commercial stage; a 128-core server-class CPU is planned to tape out in 2027 and enter mass production in 2028.
The company’s technical narrative is clear: agent workloads are characterized by a large number of concurrent sessions and frequent I/O waits, and single-core throughput is less effective than “trading wait time for another thread,” which is exactly what SMT4 is designed for. In the short term it sells IP licensing and custom development, and only in the long term will it sell chips; this also explains why it has been able to raise funding consecutively, though customer names, orders, and revenue have not been disclosed.
The risks are equally clear. The 128-core chip will not tape out until 2027 and reach mass production in 2028, and along the way it must clear three hurdles: coherent interconnect, RAS, and the software stack; moreover, the “agent CPU” it is championing is a category that has not yet been validated at scale, and whether customers are willing to pay a premium for SMT4 will only be answered once the first batch of systems ships. The signal on the capital side, however, is clear: from the 2025 GPU strategic placement to the 2026 compute-in-memory fusion and RISC-V CPU, Meituan has placed separate bets on three lines, and industrial capital is treating RISC-V server cores as a long-term track rather than a thematic investment.
Date: 2026-11-13 Source: Cyzone: Lingrui Zhixin Completes Pre-A+ Round of Several Hundred Million Yuan, Sina Finance: Meituan Leads Lingrui Zhixin Pre-A+ Round
China Business Journal: RISC-V Main Control Chip Market Grows from 1.1 Billion to 37.7 Billion Yuan in Five Years (11-14)
China Business Journal published an industry overview on November 14 offering several citable figures: by Frost & Sullivan’s measure, China’s RISC-V main control chip market grew from 1.1 billion yuan in 2021 to 37.7 billion yuan in 2025, a compound annual growth rate of 139.5%, and is projected to reach 219.4 billion yuan by 2030; the article cites Chinese Academy of Engineering academician Ni Guangnan as saying that RISC-V already accounts for 25% of the global processor market; and by the RISC-V International’s 2026 public figures, cumulative global shipments of RISC-V processor cores exceed 25 billion units. The article also outlines the foundational base of the Ruyi community: the openRuyi native operating system, the RuyiSDK development toolset, the RuyiAI software stack, the RuyiRepo software repository, and the RVCK kernel same-source project.
These figures should be used with care. The 139.5% compound growth rate comes from an extremely low base—the 1.1 billion yuan figure in 2021 was itself close to the industry’s startup phase; the 25% global share is measured in “number of processors” rather than “compute” or “revenue,” and is not the same as share in servers or phone main controls. What is genuinely informative is the structure: the article attributes growth to three lines—servers, automotive-grade, and AI edge—and explicitly identifies foundational software as the current bottleneck, consistent with what previous summits and toolchain vendors have said.
Another point worth noting is how the Ruyi community is organized. It packages the operating system, SDK, AI stack, software repository, and kernel same-source project into a single public foundation, then builds it jointly with Alibaba, ZTE, SpacemiT, Sophgo, ESWIN, and multiple distribution communities; this path of “first building the foundational pieces the industry is unwilling to build, then letting vendors plug in” is the most substantive change in China’s RISC-V software ecosystem over the past two years.
Date: 2026-11-14 Source: Eastmoney reprint of China Business Journal: RISC-V Rises, Chinese Vendors Tackle Servers, Automotive-Grade, and AI, China Business Journal original
Sophgo sophon-tools Eleven Commits: Management Plane Fixes and Default Credential Remediation (11-15)
Sophgo’s sophon-tools repository landed eleven commits between November 14 and 15, all fix-type submissions. The contents include: fixing a concurrent write race in the TimeoutMiddleware response header, retiring the old upgrade interface and correcting the SSE connection count and old stream reclamation, fixing bm_set_ip losing configuration under NetworkManager and returning success on failure, fixing DNS not taking effect and incorrect rollback, fixing symlink mishandling and upload path bypass in file management (security-related), changing CV84X6 TPU frequency reads to prefer sysfs to correct the issue of reporting only half the real value, adding X11 development headers to support GUI builds, and changing the default password for the DFSS open account and bumping to 1.11.1.
These commits have nothing to do with the instruction set, but they are critical to Sophgo’s server and accelerator card products: issues like bm_set_ip “returning success on failure” cause batch deployments to be misconfigured without warning, the SSE connection count and old stream reclamation directly affect management interface stability, and the upload path bypass and default password fall under security baselines. The TPU frequency reporting only half the value would mislead performance monitoring and tuning, and switching to prefer sysfs indicates that its value path previously relied on unreliable inputs.
From an ecosystem perspective, the quality of this kind of management software determines whether RISC-V accelerator cards can get into data centers. No matter how good the chip specs look, if out-of-band management and batch deployment tools are unstable, operations teams will not put it into production clusters. That Sophgo has continued to do fixes rather than new features during this period after National Day reflects the norm of the productization phase.
Date: 2026-11-15 Source: sophgo/sophon-tools
IV. Special Chapter on Embodied Intelligence and Robotics
Beijing’s Three-Year Embodied Intelligence Action Plan Puts RISC-V AI Compute Chips on the Breakthrough List (11-11)
The Beijing Municipal Bureau of Economy and Information Technology issued the “Beijing High-Quality Development Action Plan for the Embodied Intelligent Robot Industry (2026–2028).” The plan was finalized on November 11, and media interpretation spread widely starting November 12. In the key core technology breakthrough list under the technology innovation source project, the document names two types of chips: RISC-V-based system-level AI compute chips and high-dynamic-load N-channel MOSFETs, while also making forward-looking arrangements for ultra-low-power AI compute chips. The accompanying phrasing is “lay out whatever can be laid out, adapt whatever can be adapted”—that is, the supply side is required to produce the chips, and the demand side is required to have system integrators install them. Quantified targets include: by 2028, cultivate no fewer than 5 world-class and 10 domestically leading humanoid robot complete-machine enterprises; the Beijing-Tianjin-Hebei region to account for 25% of national complete-machine production capacity; supporting rate for key components to exceed 60%; form an annual complete-machine production capacity of 200,000 units in Beijing; and build a dedicated compute cluster with a total scale exceeding 50,000 PFlops.
Writing an instruction set architecture into the main text of a robotics-specific policy is uncommon in China. The rationale given in the plan can be reduced to three points: an open instruction set allows manufacturers to harden vector, matrix, and dedicated operators into the pipeline, adapting to on-device inference of vision-language-action models; there are no licensing fees or royalties, which is a real cost difference for robot chips with annual shipments in the hundreds of thousands and extreme sensitivity to bill-of-materials cost; and supply chain autonomy and controllability. Combined, these three make RISC-V the only general-purpose architecture named in the document.
What needs to be distinguished is policy language versus engineering reality. The plan also sets complete-machine metrics: rated load of more than 40% of self-weight, mean time between failures exceeding 2,000 hours, annual failure rate of core motion components below 10%, design life exceeding 5 years, and continuous operation endurance of more than 4 hours. These metrics constrain the power and thermal budget of on-device compute, and also explain why the document lists “ultra-low-power AI compute chips” as a separate item: compute inside a robot’s body must first survive within this power envelope.
Date: 2026-11-11 Source: NetEase reprint: Full text of the Beijing High-Quality Development Action Plan for the Embodied Intelligent Robot Industry (2026–2028), Tencent News: Beijing releases three-year action plan for the embodied intelligence industry, explicitly supporting domestic RISC-V chips
Compute Tiers for On-Device Agents: LQ7000’s TOPS and P100’s SMT4 Each Handle a Different Layer (11-15)
Looking at two domestic product lines together this week makes it clear that “agent compute” is actually tiered. Longqing Kongtian’s LQ7000 reports matrix compute of 20 to 400 TOPS, with a preinstalled large model, targeting on-device inference; Lingrui Zhixin’s P100 reports SMT4 and out-of-order width, targeting concurrent scheduling on servers and edge. The former measures how many multiply-accumulates can be done per unit time; the latter measures how many tasks a single core can hold simultaneously and whether it can still work while waiting on I/O. The two are not on the same axis.
Embodied intelligence compresses these two layers into the same power budget. The motion control layer requires hard real-time at the 50 Hz level, with only 20 milliseconds per whole-body decision, relying on deterministic response rather than peak compute; the vision-language-action layer involves inference with hundreds of millions to billions of parameters at 2.5 to 5 Hz, relying on matrix compute and memory bandwidth; the agent layer above that performs long-horizon planning and tool invocation, relying on multitasking concurrency and memory capacity. The resource demands of the three layers conflict with one another; if any one layer consumes too much power, the other two are squeezed.
This also explains why domestic manufacturers have begun using different products to target different layers: SpacemiT lines up K1, K3, and K5 as a progression from motion control to agents; Longqing Kongtian uses a scalable TOPS range to cover on-device inference; Lingrui Zhixin uses SMT4 to handle server-side concurrency. For buyers, judging whether a chip is suitable first requires asking which layer it serves, then looking at the corresponding metrics; comparing TOPS against SMT thread counts, or using motion control real-time requirements to evaluate an inference chip, will both lead to wrong conclusions.
Date: 2026-11-15 Source: Drivers.com: Longqing Kongtian LQ7000, Cyzone: Lingrui Zhixin P100 and Pre-A+ round, CCF TF 183 recap: RISC-V builds an open compute system for embodied intelligence, NetEase reprint: Beijing embodied intelligent robot industry action plan
V. Summary and Trend Observations
The clearest signal this week is that RISC-V has begun appearing in procurement decisions for “branded complete machines.” Purism switched the main controller of its next-generation Linux phone to SpacemiT K1 and publicly acknowledged that the graphics stack would have to be rebuilt from scratch; this is not a development-board crowdfunding campaign, but a manufacturer with a continuous product line changing architecture on a flagship product. K1 appears simultaneously in Start9’s router and Purism’s phone, showing that domestic RISC-V application processors can already cover both network equipment and mobile terminal complete machines.
The second thread is the combined force of capital and policy. Lingrui Zhixin raised hundreds of millions of yuan in a single round, with cumulative funding approaching one billion yuan, and its investors shifted from financial capital to industrial capital such as Meituan, JD.com, SenseTime, and SAIC; Beijing’s three-year embodied intelligence action plan puts RISC-V system-level AI compute chips on the breakthrough list. Both point to the same judgment: RISC-V’s high-value scenarios are locked into three areas—servers, automotive-grade, and embodied intelligence—and all three require long-term investment and are unsuitable as short-term themes.
The third thread is that software-side catch-up is accelerating. openRuyi enables PIE across three LLVM lines at once and brings PyTorch’s RISC-V adaptation into the distribution as a software package; ReVyOS uses a single kernel configuration to cover six platform configurations and adds post-silicon errata for K3 and EIC7700; XiangShan’s NEMU pins down the boundary conditions of virtualization and atomic operations one by one. This work does not generate launch-event material, but it determines whether development can proceed directly once the chip is in hand.
The fourth thread is that the cost of fragmentation is becoming explicit. There are at least five parallel proposals for the RISC-V matrix extension, toolchain vendors must use a unified IDE to absorb architectural differences, and functional safety still faces ten standards; the further the ecosystem moves toward high-value scenarios, the higher the value of standard convergence. This week, two articles from IAR and FPRox said the same thing from different angles: the next stage of RISC-V competition is not in single-core benchmarks, but in “who defines the common baseline.”
Finally, the timelines. Purism points to delivery in 2028, FIORA5 points to a prototype in 2027 and licensing in 2029, Lingrui Zhixin points to tape-out in 2027 and mass production in 2028, and Longqing Kongtian has not given a mass production node. The fulfillment period for this batch of products and chips is generally two years out; during the validation window in between, problems in any link of the software stack, certification, or supply chain will delay deployment. For observers, three points worth watching next are: whether the matrix extension draft can converge, whether the K1 phone’s graphics stack can run on test devices in 2027, and the first tape-out results of 128-core-class RISC-V server chips.
Appendix
Source Verification Table
| Source | Verification Result |
|---|---|
| Google News aggregation (24 Chinese and English query sets, 72-hour window) | Substantive new items within the window: Purism Librem 5 v2, Lingrui Zhixin Pre-A+ funding, Longqing Kongtian LQ7000, FIORA5 crowdfunding, VVCE tape-out, IAR summit recap, China Business Network industry overview. Excluded items: the Milk-V query returned 16 items about milk and dairy products; popular-science pieces such as RISC-V Explained and NoobFeed; analyses of Arm’s licensing mechanism; Tendrils Compute’s $30 million funding (interconnect network architecture, not RISC-V); BSC’s SC26 overview of RISC-V and quantum computing used only as background, not listed separately |
| Hacker News / Phoronix / CNX Software / LinuxGizmos / RISC-V International | HN returned one item on FPRox’s RISC-V matrix extensions, included; Phoronix’s RISC-V channel had no dedicated new articles within the window, and the two items on QEMU memory tiering and Redox OS are not directly related to RISC-V, so not included; CNX returned one item on a desktop OS for ESP32-P4, included; LinuxGizmos’s lead story was Librem 5 v2; the riscv.org blog had no new posts within the window |
| GitHub organization and key repository checks (30+ organizations and repositories) | Active within the window: openRuyi 20 commits, revyos/linux 20 commits, sophgo/sophon-tools 11 commits; on the XiangShan side, NEMU 4 commits, XiangShan 2 commits, difftest 1 commit, XSCache 1 commit. DAMO XuanTie (XUANTIE-RV), ESWIN (eswincomputing), Nuclei (Nuclei-Software), and openhwfoundation had zero pushes within the window. The anonymous API hit the 60-requests-per-hour rate limit, so organization-level activity was checked instead via each repository’s Atom commit feed and organization pages; pushes to non-default branches may be undercounted |
| riscv100.top inspection | No changes: the homepage What’s New, Upcoming, and committee members, as well as the six SPEC CPU2006/2017/2026 leaderboards under the cross-architecture and RISC-V channels, showed no new records or score changes, so no space is devoted to them in the body |
| RISC-V member website inspection (391 sites) | Took 237 seconds; 218 unchanged, 100 flagged as changed, 73 failed to fetch (including ISCAS, Beijing Institute of Open Source Chip, Tencent, NVIDIA, Synopsys, ADI, DENSO, and others). The changed items, checked one by one, were all homepage announcement rotations or marketing updates (SiFive, AMD, Canonical, Codasip, imec, and several universities), with no new products, funding, or architecture launches, so not included |
| openRuyi official news page and main repository | The news page’s latest items remain the two from 2026-10 (ecosystem infrastructure, Ruyi Firmware) and the 2026.09 release; no new version within the window; the 20 commits in the main repository have been logged and included |
Full Source List
- [1] Purism: Librem 5 v2 crowdfunding page — https://puri.sm/librem-5-v2-crowdfunding-campaign
- [2] LinuxGizmos: Librem 5 v2 adopts eight-core RISC-V processor — https://linuxgizmos.com/librem-5-v2-linux-smartphone-adopts-eight-core-risc-v-processor
- [3] OMG! Ubuntu: Purism unveils second-generation Librem 5 — https://www.omgubuntu.co.uk/2026/11/librem-5-risc-v-new-linux-phone
- [4] FPRox: RISC-V Matrix Extensions — https://fprox.substack.com/p/risc-v-matrix-extensions
- [5] FURSOY/fursoy-cpu — https://github.com/FURSOY/fursoy-cpu
- [6] LavX News: FURSOY open-source core boots Linux and closes at 500MHz — https://news.lavx.hu/article/fursoy-publishes-open-source-risc-v-core-that-boots-linux-and-hits-500-mhz-in-synthesis
- [7] openRuyi-Project/openRuyi — https://github.com/openRuyi-Project/openRuyi
- [8] revyos/linux — https://github.com/revyos/linux
- [9] OpenXiangShan/NEMU — https://github.com/OpenXiangShan/NEMU
- [10] OpenXiangShan/XiangShan — https://github.com/OpenXiangShan/XiangShan
- [11] OpenXiangShan/XSCache — https://github.com/OpenXiangShan/XSCache
- [12] EET China: RISC-V China Summit recap — one toolchain addresses architectural fragmentation and functional safety challenges — https://www.eet-china.com/newsexpress/75861.html
- [13] EE Times: Inside Architect Labs’ Two-Week Chip Design — https://www.eetimes.com/inside-architect-labs-two-week-chip-design
- [14] CNX Software: Desktop operating systems on ESP32-P4 — https://www.cnx-software.com/2026/11/14/desktop-operating-systems-for-esp32-p4-based-pcs-meet-sylphide-os-and-bare-metal-do-stuff-bmds/
- [15] MyDrivers: Longqing Kongtian launches edge AI compute chip LQ7000 — https://news.mydrivers.com/1/1165/1165672.htm
- [16] Eastmoney: Longqing Kongtian launches LQ7000 — https://emh5wap.eastmoney.com/Info/Detail/202611133930718593
- [17] Foro3D: FIORA5 opens funding round on Growceanu — https://foro3d.com/en/2026/noviembre/fiora5-abre-ronda-en-growceanu-para-su-procesador-risc-v-de-ia-en-el-b.html
- [18] ITKeyMedia: Vest Ventures invests in FIORA5 — https://itkey.media/west-romanian-vest-ventures-backs-fiora5-as-it-moves-toward-its-first-processor-poc
- [19] The Hindu: VVCE achieves milestone in indigenous chip design — https://www.thehindu.com/news/national/karnataka/vvce-achieves-milestone-in-indigenous-chip-design/article71704681.ece
- [20] Cyzone: Lingrui Zhixin completes several-hundred-million-yuan Pre-A+ round — https://www.cyzone.cn/article/853804.html
- [21] Sina Finance: Meituan leads Lingrui Zhixin Pre-A+ round — https://finance.sina.com.cn/stock/t/2026-11-13/doc-iniyqnee8028075.shtml
- [22] Eastmoney repost of China Business Network: RISC-V rises as Chinese vendors tackle servers, automotive-grade, and AI — https://finance.eastmoney.com/a/202611143931312434.html
- [23] China Business Network original article — http://www.cb.com.cn/index/show/bzyc/cv/cv135295061646
- [24] sophgo/sophon-tools — https://github.com/sophgo/sophon-tools
- [25] NetEase repost: Beijing Embodied Intelligence Robotics Industry High-Quality Development Action Plan (2026–2028) — https://www.163.com/dy/article/LBGCQ6G20552IBJ1.html
- [26] Tencent News: Beijing releases three-year action plan for embodied intelligence industry — https://news.qq.com/rain/a/20261112A0BLEI00
- [27] CCF TF 183 recap: From chips to robot applications, RISC-V builds an open-source compute system for embodied intelligence — https://zhuanlan.zhihu.com/p/2081684550313325606
Limitations: This is the Monday tri-daily edition, with a window covering 72 hours from 05:00 on November 13 to 05:00 on November 16. The GitHub anonymous API hit its 60-requests-per-hour rate limit during this edition, so organization and repository activity was checked instead via each repository’s Atom commit feed and organization pages; pushes to non-default branches may therefore be undercounted. All five RISC-V matrix extension proposals remain at the draft stage; the article’s wording follows the authors’ claims, and the specifications may change at any time. Longqing Kongtian’s 20–400 TOPS range is not mapped to specific models or power consumption, and the 90% figure for in-orbit scenarios lacks third-party verification; FIORA5’s area and energy data and Architect Labs’ acceleration and power comparisons are vendor claims lacking independent benchmarks. Purism’s crowdfunding project carries risks of delivery delays and unmet goals, and the $1,599 figure is the crowdfunding price rather than the retail price. Beijing’s embodied intelligence action plan was finalized on November 11, a policy document predating this window; it is included here as background and its finalization date is noted in the index.