RISC-V Daily Intelligence Report (2026-11-12)
Research window: 2026-11-11 05:00 ~ 2026-11-12 05:00 (Beijing time, 24 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/cvfpu, DAMO XuanTie, ESWIN, Sophgo, Nuclei, Espressif, SpacemiT and spacemit-robotics, Sipeed, revyos, openRuyi-Project, and Radxa Linkr.
Index for This Issue
- Today’s Highlights: Beijing issues three-year action plan for embodied intelligent robots; 200,000-unit annual capacity and 50,000 PFlops cluster quantify demand side (11-11)
- I. Open Source Community and Software Ecosystem
- RevyOS kernel: 101 commits in a single window: K3 display and USB3 land, EIC7700 fan and thermal zones, SG2044 SRD3 dual-board merged (11-11)
- openRuyi 36 commits: OpenTelemetry and four Python certification package security updates, three OpenCL packages merged, kernel upgraded to 7.3.4/6.18.63 (11-11)
- Espressif esp-riscv-trace 0.2.0: RISC-V instruction-level trace adds four commands: decode/profile/stats/history (11-11)
- ch32rv-clang: CH32 Arduino core switches to RISC-V-only clang and lld, builds only the RISCV backend (11-11)
- NEMU four commits: SMBLOCKCTL adds PF refresh bit, RVV zero-length access sets VS dirty, AME matrix fast path (11-11)
- openRuyi-Project creates ruyi-deck: an operations console for repositories and releases (11-11)
- II. Product and Ecosystem Developments
- T-Head spin-off moves from rumor to announcement: Alibaba’s 11-10 evening announcement of proposed Hong Kong listing spin-off; HKEX confirms separability (11-11)
- III. Deep Dive on Domestic Teams
- XiangShan Kunminghu V3: 18 commits in one TLB refactor; permission checks and PTW bypass converge to a single exit (11-11)
- SpacemiT llama.cpp: F16 GEMM main loop switched to double buffering; attention initialization and ASR scheduling fixes (11-11)
- SpacemiT ai-sdk and model-zoo-vision: inference threads now allocated by TCM core budget (11-11)
- Sophgo 19 repos updated weekly: cviruntime ships 2026.11.12 weekly version, full SDK suite aligned in same batch (11-11)
- Nuclei tool guide: GDB real-time variable watch added to 2026.10 changelog (11-11)
- Sipeed SLogicWeb 14 commits: logic analyzer frontend adds behavior settings and cache governance (11-11)
- IV. Special Chapter on Embodied Intelligence and Robotics
- SpacemiT SuperMimic expands from LingLong to G1: standby plus four optional actions and a general action catalog land in the same batch (11-11)
- V. Summary and Trend Observations
- Appendix: Complete Source List and Source Verification Table
Today’s Focus: Beijing Issues Three-Year Action Plan for Embodied AI Robots, Quantifying Demand-Side Targets with 200,000 Units Annual Capacity and a 50,000 PFlops Cluster (11-11)
The Beijing Municipal Bureau of Economy and Information Technology issued the “Beijing Embodied AI Robot Industry High-Quality Development Action Plan (2026–2028)” on November 11, effective from the date of issuance through December 31, 2028. The document replaces the directional language of earlier policy rounds with assessable numbers: by 2028, cultivate no fewer than 5 world-class and 10 domestically leading humanoid robot OEMs; have the Beijing-Tianjin-Hebei region account for 25% of national OEM production capacity, with a key component localization rate above 60%; build 200,000 units/year of OEM production capacity in Beijing; and add a dedicated embodied AI computing cluster with total scale exceeding 50,000 PFlops, while promoting “pairing” arrangements between key enterprises and computing-node builders to create supply-demand coordination for compute.
The technology roadmap is more specific than before. The plan explicitly supports multi-route layout and integration of vision-language-action models (VLA), world models (WM), reinforcement learning, and others, and splits the research push into “brain” and “cerebellum” sides: the brain side covers multi-source information fusion, physical-space understanding, and human emotion recognition, with the goal of enabling robots to understand physical-law constraints and proactively recognize non-verbal commands such as human actions and facial expressions; the cerebellum side covers real-time inference, cross-embodiment transfer, and self-learning evolution, targeting multi-task generalization, highly reliable execution, and ultra-precision manipulation. On the manufacturing side, the plan sets out hard targets one by one: build 3 to 5 world-class OEM platforms; achieve a rated payload of at least 40% of the robot’s own weight; mean time between failures exceeding 2,000 hours; an annual failure rate below 10% for core motion components and end effectors; a design service life exceeding 5 years; battery endurance of more than 4 hours under continuous operation; and reduce the cost of a single high-DOF dexterous hand to under RMB 10,000.
Governance and capital arrangements are also fully addressed. The plan proposes a “one machine, one code” full-process management system covering R&D, production, scrapping, and recycling; establishes a “one-on-one” service mechanism for OEM listings; cultivates no fewer than 200 innovative enterprises in areas such as complete machines, key components, embodied AI models, and data collection; promotes the establishment of 100 robotics research teams at Beijing universities and research institutes; and explicitly opposes “involution-style” competition. These targets are not arbitrary: on October 19, MIIT disclosed that in the first half of 2026, domestic humanoid robot OEM products exceeded 400 models with output above 40,000 units, and full-year output is expected to exceed 100,000 units; an October report from CAICT presented the other side—among publicly disclosed winning bids in the first half, education and research institutions accounted for about 55.5%, government and public platforms about 20.6%, and industrial and technology enterprises about 21.1%, with a considerable share of equipment averaging fewer than 10 powered-on days per month, and complete-machine prices possibly accounting for only 40% to 60% of total cost of ownership.
For the RISC-V line, the value of this plan lies in turning the “demand side” from a vision into procurement parameters. The 200,000 units/year OEM capacity, the 60% local key-component matching rate, and dexterous hands under RMB 10,000 together point to the cost and reliability of chips at two levels: joint control and onboard compute. The joint side requires hard real-time performance, low power, and deterministic networking; the onboard side requires compute and memory bandwidth capable of running VLA inference—and both must be mass-produced under the constraint of 2,000 hours without failure. Among domestic RISC-V suppliers currently able to address these two levels are SpacemiT’s K3 series (onboard compute), Sophgo’s SG and BM series (server and edge), Nuclei’s N1000/NI1000 (real-time cores and vector-matrix cores), and MIPS’s Atlas M8500 (joint-level real-time core; its Actus platform was reported on 11-09). The 50,000 PFlops dedicated cluster falls on the training and data side and is directly related to accelerator cards such as PPU and DCU. It should be noted that the plan itself does not name any instruction set or supplier; the above correspondences are a comparison against the current state of the supply side, not a policy designation.
Date: 2026-11-11 Source: Wallstreetcn: Beijing Embodied AI Three-Year Plan Released, Beijing Business Today: Beijing Robotics Industry Plan Unveiled
I. Open-Source Community and Software Ecosystem
RevyOS Kernel: 101 Commits in a Single Window — K3 Display and USB3 Land, EIC7700 Fan and Thermal Zones, SG2044 SRD3 Dual Boards Merge (11-11)
The revyos/7.3.y branch of the revyos/linux repository landed 101 commits in this window, split into three batches by time. The largest batch centers on 11-11 13:06 (Beijing time), mainly platform features and FROMLIST backports: on the SpacemiT side, the K3 GPU node and display power domain, Type-C Port A on the K3 Pico ITX, USB3 PHY and the FUSB301 autonomous controller, PCIe asserting PERST# only after PHY initialization, UFS WriteBooster support, and disabling amdgpu’s dynamic PCIe link-speed switching on K3; on the ESWIN side, the EIC7700 gains an hwmon fan controller, PWM, thermal zone nodes and CPU/device frequency scaling with a 1.4GHz maximum; the SiFive HiFive Premier P550 fills in USB, SATA, PCIe, eMMC and SD; on the T-Head side, the Lichee Pi 4A keeps USB powered during handover and USB is enabled for the Milk-V Meles.
Beyond the chip platforms there is a layer of generic changes. A batch of FROMLIST patches around RISC-V page-table accessors was backported wholesale: routing all page-table access through accessor functions, allowing non-idempotent accessors, fixing the memory type used when writing page tables, displaying the N and MT bits in ptdump according to their enable state, exposing all page-table bits to assembly, adding a three-argument macro and a replaceable link register to alternative, and fixing the selection logic for DMA_DIRECT_REMAP. On the KVM side, all counters are exposed to VS-mode for bootloader compatibility, and the XLINUXENVCFG cpufeature guards SENVCFG access.
The third batch is the 14:52 defconfig consolidation: expanded from the six commits reported yesterday into a full matrix — K1, K3, DP1000, EIC7700, TH1520 and SG2042/SG2044 enabled one by one, plus Picoheart errata, common options and debian_defconfig, along with linuxboot configuration, both cross and native CI builds, and packaging-side HACKs such as “use the major version number as the package name.” Taken together, the point of this day is not any single new feature but consolidating board support previously scattered across branches into one combination that compiles and produces packages.
Worth a separate look is the SG2044 board line: the device trees for the SRD3-10-LB and SRD3-12 boards both merge, with SRD3-12 also carrying a large-BAR variant, plus a second C2C1 x4 PCIe port, an eFUSE controller, and PCIe windows aligned with the EDK2 implementation. For teams building RISC-V servers, board descriptions matching the firmware’s conventions matters more than one more driver — inconsistent window partitioning shows up directly as incomplete device enumeration in multi-card scenarios.
Date: 2026-11-11 Source: revyos/linux
openRuyi: 36 Commits — OpenTelemetry and Four Python Auth Packages Get Security Updates, Three OpenCL Packages Merge, Kernel Bumped to 7.3.4/6.18.63 (11-11)
The openRuyi main repository saw 36 commits in this window, closing out two supply chains at once. The first is observability: the python-opentelemetry family is bumped to 1.45.2, with the api package explicitly marked security-fixes, semantic-conventions bumped to 0.66b2, and two new subpackages, exporter-otlp-common and exporter-http-transport, effectively completing the optional OTLP export path. The second is authentication: python-pyjwt 2.15.1, python-oauthlib 4.0.1, python-joserfc 1.8.0 and python-authlib 1.9.3 were all updated the same day under the security-fixes label. Both fall into the category of “won’t make headlines but determines whether you can ship.”
The kernel saw two version bumps the same day: linux to 7.3.4 and linux-lts to 6.18.63, both with patchset and config refreshes. The companion kernel-team-tools repository pushed tags to v7.3.4-2 and v6.18.63-3, registered 7.0.14-9/10 and 7.2.15-4 in the manifest, and restricted the XTHEADVECTOR and XLINUXMEMALIAS options to rva20 platforms — an easily overlooked change with broad impact: the vector and memory-alias options are no longer enabled by default on higher-baseline platforms, avoiding behavioral drift between different RVA baselines.
On new packages, three OpenCL packages merged the same day: opencl-headers, opencl-icd-loader and clinfo; also fwupd, qemu-rustsbi and a batch of Rust dependencies (rust-libc 0.2.185, memchr 2.7.6, hashbrown 0.17.1, clap-verbosity-flag 3, rustc-literal-escaper 0.0.7). qemu-rustsbi is worth noting — a Rust-written SBI reference implementation entering a distro repository means multilingual implementations on the RISC-V firmware side are starting to be taken seriously at the packaging level; clinfo, meanwhile, turns “are the OpenCL devices on this machine actually enumerated” into a command you can verify directly.
The rest is routine maintenance: cmake enables ccmake, htop 3.5.3, nano 9.3, openldap 2.7.2 (restoring tests and securely initializing slapd), cyrus-sasl fixing GCC 15 compatibility with patch provenance noted, openjdk 17/21/25 refreshing the CA enhancement patch, clang-wrap bumped to 20261111, and go-github’s discordgo forked into a standalone package declaring a conflict with the original.
Date: 2026-11-11 Source: openRuyi-Project/openRuyi, kernel-team-tools
Espressif esp-riscv-trace 0.2.0: RISC-V Instruction-Level Trace Adds decode/profile/stats/history Commands (11-11)
Espressif pushed esp-riscv-trace to 0.2.0. The tool decodes hardware instruction traces captured on Espressif RISC-V chips: it reads .esptrace captures produced by OpenOCD, or core ELFs generated by esp-coredump with trace annotations, reconstructs the program execution flow, resolves each PC to a function and source line, then outputs a text dump, a hotspot table, the call history leading up to a crash, or call stacks ready to feed into a flame graph; points where continuity is lost are marked as gaps rather than guessed and filled in. The four commands added in 0.2.0 push this capability from “can capture” to “can analyze”: decode lists executed instructions, profile writes flame-graph stacks, stats counts instructions by source line, and history, together with the open_capture interface, opens up merging and reviewing multiple captures.
The supported chip list covers ESP32-C5, C6, C61, H2, H21, H4, P4 and S31, with package formats spanning two generations, Processor Trace 1.0 and Efficient Trace 2.0. Usage is to enable trace configuration in ESP-IDF, capture with OpenOCD’s esp riscv_trace start/stop, and after a panic the firmware stops trace before the debugger halts, avoiding loss of the final segment.
The significance here is “debug visibility on the RISC-V device side.” Performance analysis on sub-dollar RISC-V MCUs used to rely on instrumentation and counters; instruction-level trace was long a feature of high-end debuggers. Espressif has turned it into a pip-installable Python tool and opened up the interface for reading captures, effectively making “how many instructions did this code actually execute, and which line is the hotspot” a routine technique. For teams doing inference operators and real-time control on RISC-V, the marginal value of such tools often exceeds another bump in clock speed.
Date: 2026-11-11 Source: espressif/esp-riscv-trace
ch32rv-clang: CH32 Arduino Core Switches to RISC-V-only clang and lld, Building Only the RISCV Backend (11-11)
The community organization CH32 RISC-V User Group released ch32rv-clang: clang and lld built from upstream LLVM releases with only the RISCV backend retained, packaged as one archive per host system for installation by ArduinoCore-CH32RV through the Arduino Board Manager. This release deliberately subtracts — the archive contains no C library and no compiler runtime, and still links against newlib-nano and libgcc from the xPack GNU RISC-V Embedded GCC for now; in exchange, size and the install path are simplified: a single root directory, 12 tools under bin (a single multicall binary plus symlinks on Linux and macOS, separate executables on Windows), ready to use after extraction.
On the ISA side it supports WCH QingKe’s XW compressed instructions, corresponding to the Xwchc extension in LLVM 19 and later (with _xwchc in -march). Which cores have XW is managed by the core’s device table — and this is stated quite clearly: on cores without XW such as QingKe V3A and V4A, the program halts as soon as the first such instruction executes, so you cannot guess from the chip model. Prebuilt binaries cover five hosts: x86_64 and aarch64 Linux, x86_64 and aarch64 macOS, and Windows MSVC. Version numbers follow “LLVM version-build number,” e.g. v23.1.5-1; one tag triggers builds for all hosts and publishes artifacts with .sha256. The build scripts are MIT, and the bundled LLVM is Apache 2.0 with LLVM Exceptions.
For this product line, switching compilers is not the goal — lowering the barrier to entry is. RISC-V MCUs in the ten-cent class such as the CH32V003 previously relied on GCC toolchains in the Arduino ecosystem, with fragile versions and install steps; plugging LLVM tarballs into the Board Manager removes half of the most common deterrent: “can’t install the toolchain.”
Date: 2026-11-11 Source: Adafruit: A RISC-V-only clang + lld for the CH32 Arduino core, ch32-riscv-ug/ch32rv-clang
NEMU: Four Commits — SMBLOCKCTL Gains PF Flush Bit, RVV Zero-Length Accesses Set VS Dirty, AME Matrix Fast Path (11-11)
NEMU landed four commits in this window. On the feature side, SMBLOCKCTL gained a PF flush enable bit (#1285), turning this control bit from a reserved bit into configurable behavior; on the RVV side, a fix for zero-length gather, slide and mask-class operations failing to set VS dirty (#1286) — a missed vector-state dirty bit in virtualization scenarios causes the host to misjudge whether context needs saving, a classic defect that only surfaces under specific instruction combinations. On the MMU side, a fix for reporting the next-page base address when the second page has a non-canonical address.
One performance commit deserves more attention: enabling the matrix load/store fast path in the shared ref (#1284). AME is the XiangShan-family matrix extension, and over the past year NEMU’s support for it has moved from completing instruction semantics to optimizing the memory-access path, indicating the matrix unit now runs enough in the simulator that performance has become the bottleneck. The difftest side had two coordinating commits the same day: the nightly workflow dropping the Firtool distribution’s latest tag (#1040), and vectorizing MMA’s CPU check while reusing already-decoded operands (#1038).
Read together, the center of gravity of the verification infrastructure is shifting from “is it functionally correct” to “does it run fast enough.” The state volume and timing complexity introduced by the matrix extension make the throughput of the pure-software reference model the determining factor in regression duration; the performance fast path and check vectorization are two moves in the same direction.
Date: 2026-11-11 Source: OpenXiangShan/NEMU, OpenXiangShan/difftest
openRuyi-Project Creates ruyi-deck: An Operations Deck for Repositories and Releases (11-11)
openRuyi-Project created the repository ruyi-deck within the window; the commit message is just “Init repo,” and the README gives a one-line positioning: The Operations Deck for openRuyi Repositories and Releases, under the Mulan Permissive Software License v2. From the name and description, this is a standalone component splitting “the operational surface of repositories and releases” out of the main repository, aimed at operations actions such as builds, releases and repository state management.
The news itself is small, but the placement is key. Over the past year the openRuyi main repository has taken on too many roles: it is both the SPECS packaging repo and the home of configs and patches, and tests and CI keep being added to it. As the release cadence accelerates to the point of needing weekly tags (the density of tags like 7.3.4-2 and 6.18.63-3 in kernel-team-tools is a reference), splitting operations logic into a component is an inevitable step. For downstream distributions and mirror sites, what to watch next is whether this repo will pin down the definitions of the three things: “release artifacts, signing, channels.”
Date: 2026-11-11 Source: openRuyi-Project/ruyi-deck
II. Product and Ecosystem Developments
T-Head spin-off moves from rumor to announcement: Alibaba’s evening announcement on 11-10 proposes a Hong Kong listing, HKEX confirms the spin-off is permissible (11-11)
Alibaba issued an announcement on the evening of November 10 proposing to spin off its wholly-owned subsidiary T-Head for a separate listing on the Main Board of the Hong Kong Stock Exchange; HKEX has confirmed that Alibaba may proceed with this spin-off; T-Head had already submitted its A1 listing application confidentially as early as September 11. The blanks left in the announcement are equally important: the offering size and the extent of the reduction in Alibaba’s shareholding are both undetermined, and because the highest applicable percentage ratio is below 5%, this spin-off does not even constitute a disclosable transaction for Alibaba—using a transaction that is barely material to its own financial statements to bring its hardest technology asset to the doorstep of the public market.
The pricing range remains the most contentious item. Previously reported fundraising size was US$2 billion to US$3 billion; on valuation, a professional investment institution source gave US$20 billion to US$30 billion, while JPMorgan’s rough benchmarking estimate in January this year was US$25 billion to US$62 billion. For reference, the total market capitalizations of the already-listed Tianshu Zhixin and Biren Technology in Hong Kong both exceed HK$50 billion. The supporting business data is: as of April 2026, the Zhenwu series AI chips have shipped a cumulative 560,000 units, serving more than 650 enterprise customers, covering more than 20 industries, with external commercial customers accounting for more than 60%, and deployment scale in the financial sector exceeding 100,000 cards, covering more than 150 financial institutions; the Zhenwu 810E (PPU) features 96GB HBM2e, 700GB/s inter-chip bandwidth, and power consumption below 400W, while the V900 released in September has 216GB memory, 1200GB/s inter-chip bandwidth, supports FP8 and FP4, can connect up to 500,000 chips to form a cluster, and is scheduled for mass production in Q1 2027.
The relationship with RISC-V needs to be restated, because this framing has been repeated by the media for three rounds: the listing entity this time is T-Head, and its product focus is PPU (AI accelerator cards) and Yitian (Arm server CPU); the XuanTie RISC-V processor IP was spun off from T-Head in 2024 and transferred to DAMO Academy, which only does RISC-V processors. When media inventory T-Head’s product matrix, they often include “XuanTie RISC-V IP cumulative shipments exceeding 6 billion units” as well, but that is a group-level historical framing and does not mean the listed assets include the RISC-V IP business. This publication made the same distinction in its 11-10 and 11-11 issues; this issue records only the two new facts: the announcement and the pricing range.
What truly determines valuation is not the narrative but two splits in the prospectus: of the tens-of-billions-level annualized revenue, how much comes from supplying Alibaba Cloud versus direct procurement outside the group; and whether chips deployed in Alibaba Cloud’s ten-thousand-card clusters are settled at fair market prices or allocated internally. The former determines whether the market prices it as an “independent chip company” or an “Alibaba Cloud cost center,” while the latter determines the quality of gross margin after separation from the parent. A small-percentage offering that retains control after the spin-off is essentially about first obtaining a public-market pricing anchor and preserving room for subsequent capital operations.
Date: 2026-11-11 Source: 36Kr: After burning 190 billion, Alibaba pushes T-Head to the market, East Money reposting Time Weekly
III. Deep Dive into Domestic Teams
XiangShan Kunminghu V3: 18 TLB Refactoring Commits in One Pass, Permission Checks and PTW Bypass Consolidated to a Single Exit (11-11)
The kunminghu-v3 branch of XiangShan landed 21 commits in this window, 18 of which are a refactor(TLB) series with the same day and same timestamp. None of these 18 commits changes behavior; all are structural governance: naming the fields of TLB read results and PTW bypass results, sharing dynamic VPN index selection, extracting S0 address preparation and pre-checks into independent steps, reusing the TLB storage query and response interfaces, extracting permission and PBMT checks into the TLB source, assigning permission results before pre-fault override, aggregating access check information by port, reusing the permission bundle and removing the unused G-stage valid bit.
The subtraction is just as thorough: removing unused blocking TLB support, TLB storage scaffolding, the saveLevel configuration, the partialStaticPMP parameter and L1-side helper functions; consolidating the TLB refill connection into one, centralizing translation-stage selection in one place, and simplifying TLBFA hit response selection. There are also three non-TLB commits: fixing CMOUnit channel selection in MissQueue (#6907), reorganizing ROB bundle organization and clarifying IO grouping (#6881), and adding an ISA smoke test to CI’s EMU Basics (#6805).
A single batch of 18 pure refactoring commits usually means the module is stable enough to “change only structure, not behavior”—modules still changing rapidly in functionality cannot do this, because refactoring and bug fixing would be mixed together and impossible to bisect. TLB and PTW are the hardest-to-debug part of the memory access path; once permission checks and bypass results are consolidated to a single exit, the change surface for adding extensions like Svnapot and Svpbmt later will be much smaller. Adding an ISA smoke test on the CI side is the other half of the same logic: structure can change only if the regression surface is dense enough.
Date: 2026-11-11 Source: OpenXiangShan/XiangShan
SpacemiT llama.cpp: F16 GEMM Main Loop Switched to Double Buffering, Attention Initialization and ASR Scheduling Fixed (11-11)
The spacemit-com llama.cpp repository (mtmd-backend branch) landed two commits. The first changes the F16 GEMM main loop in the ggml-spacemit backend to double buffering (#62): each iteration of the matrix multiply was waiting for the previous block of data to finish loading, and double buffering overlaps the loading of the next block with the computation of the current block—a typical optimization for memory-bound scenarios. The second fixes attention initialization and improves scheduling of the ASR path (#63)—the latter shows that this backend serves more than text generation; speech recognition also runs on the same runtime.
The two commits differ in nature but point to the same thing: inference performance on K-series chips has entered the “tune the main loop” stage. Commits in previous months focused on wiring up operators correctly and completing the quantization path; now pipeline-level overlap and scheduling are being addressed. For teams using this backend, double buffering on the F16 main path means visible throughput changes for non-quantized models, while the attention initialization fix matters more for correctness—improper initialization often manifests in multi-turn dialogue and long contexts as answer drift rather than a crash, which is harder to locate.
Date: 2026-11-11 Source: spacemit-com/llama.cpp
SpacemiT ai-sdk and model-zoo-vision: Inference Threads Now Allocated by TCM Core Budget (11-11)
SpacemiT’s ai-sdk and model-zoo-vision repositories were updated on the same day. ai-sdk merged the change adding the file2md component (#12) and synchronized the avatar and vision submodules; the key commit in model-zoo-vision changes the thread count of the inference execution provider to be allocated by TCM core budget (#117)—previously the thread count was probably set by CPU core count or a fixed value, but now available threads are derived from the tightly coupled memory budget, preventing inference threads from competing with real-time tasks for the same on-chip memory.
This change looks like a parameter adjustment but is actually a common pitfall in edge deployment: TCM is a limited, per-core resource, and too many inference threads will squeeze the on-chip space of real-time control or display paths, with symptoms of occasional stutter rather than errors. Binding thread count to the TCM budget effectively writes this implicit constraint into the configuration layer.
The addition of the file2md component reveals the SDK’s direction: bringing “AI application preprocessing” such as document-to-Markdown conversion into unified components, rather than having each application integrate a parser separately. For teams building on-device document QA and multimodal applications, the value of such components lies in reducing repeated integration.
Date: 2026-11-11 Source: spacemit-com/ai-sdk, spacemit-com/model-zoo-vision
Sophgo 19 Repos Weekly Update: cviruntime Tags 2026.11.12 Weekly Release, SDK Suite Aligned in the Same Batch (11-11)
At the end of the window, Sophgo pushed to 19 repositories simultaneously within seven minutes: sophpi, cviruntime, cvikernel, osal, SensorSupportList, rt-thread, u-boot-2021.10, ramdisk, osdrv, opensbi, linux_5.10, ive, isp-tool-daemon, isp_tuning, fsbl, cvi_rtsp, cvi_alios_open, cvi_mpi and build. The cviruntime commit message directly says weekly rls 2026.11.12, indicating this is a weekly-aligned release train rather than 19 independent feature commits.
This “whole-suite same-batch” pattern has persisted at Sophgo for a long time, and its value to downstream is version composability: bootloader (fsbl), firmware (opensbi), kernel (linux_5.10), runtime (cviruntime), operator library (cvikernel), multimedia (cvi_mpi, ive, cvi_rtsp) and build system (build) align to the same version number on the same day, so the SDK customers receive will not have runtime and operator library version mismatches. The cost is that individual repositories’ commit histories are flattened by the release cadence, making it hard for outsiders to see what specifically changed from the commits—details require reading the release notes.
For teams building edge AI appliances, what is worth noting here is the cadence itself: weekly updates mean the cycle from problem feedback to fix is measured in weeks, and this cycle should be factored into project scheduling during selection rather than expecting the open-source community’s “PR anytime” model.
Date: 2026-11-11 Source: sophgo organization repositories
Nuclei Tool Guide: GDB Live Variable Watch Added to 2026.10 Changelog (11-11)
Nuclei’s nuclei-tool-guide repository landed two commits, both documentation-side: adding the GDB live variable watch feature to the 2026.10 changelog and supplementing usage instructions in non-stop mode (#90). Live watch means periodically reading and refreshing variable values while the target is not halted, used to observe continuously changing quantities such as counters, state machines and sensor readings—the traditional approach is to set a breakpoint, halt and read, which greatly disturbs real-time control loops.
This is listed separately because debugging experience has long been a weak spot in the RISC-V MCU ecosystem: compilers and debuggers exist, but the degree of “convenience” still lags mature commercial toolchains. Nuclei writing such features into monthly changelogs shows the toolchain team is improving usability at a perceptible cadence, rather than only bumping version numbers when new cores are released.
For teams building products with Nuclei cores, the practical significance of such updates is the time cost during debugging. Once variable observation in non-stop mode is stable, many problems that previously required serial printing to diagnose can be viewed directly in GDB.
Date: 2026-11-11 Source: Nuclei-Software/nuclei-tool-guide
Sipeed SLogicWeb 14 Commits: Logic Analyzer Frontend Adds Behavior Settings and Cache Governance (11-11)
Sipeed’s SLogicWeb repository landed 14 commits in the window, all around the logic analyzer’s web frontend. On the feature side, it adds a “Behavior” settings section and corresponding i18n labels, and adds three developer paths: clear cache, clear application data, and reset saved data; on the stability side, it fixes a set of issues related to Loop capture: throttling display scanning by the Loop sink, replacing rather than appending entire segments for Loop windows, maintaining trigger state under slow signals and delayed scanning, ensuring the board keeps supplying data when the sink falls behind, and governing consumer lag rather than waiting for the entire backlog.
The technical core of this commit set is “no data loss during long captures.” The value of a logic analyzer lies in catching intermittent timing problems, and the premise of intermittent problems is often continuous capture for tens of minutes or even hours; once frontend rendering cannot keep up with the sample rate, frame drops or memory blowups occur on the browser side. Replacing windows, throttling scanning to consumption speed, and propagating backpressure from the frontend to the device side are hurdles such tools must clear.
In the same window, Sipeed’s TangMega-138K-example also merged two PCIe-related changes: replaying SerDes startup registers after REFCLK stabilizes, and supplementing DDR3 notes for the 138K console. The former is a typical fix for “link fails to come up due to uncertain power-on timing,” an FPGA-side hardware adaptation lesson.
Date: 2026-11-11 Source: sipeed/SLogicWeb, sipeed/TangMega-138K-example
IV. Embodied Intelligence and Robotics Special Chapter
SpacemiT SuperMimic Expands from LingLong to G1: Standby Plus Four Optional Actions and Common Action Catalog Landed in the Same Batch (11-11)
Two spacemit-robotics repositories landed SuperMimic extensions on the same day: humanoid_unitree_g1 adds SuperMimic action mode (#16), configuring a standby state and four optional actions, sharing one CPU inference scheduler, and supplementing cancellation timing, operator binding, audio prompts, resource preparation and usage documentation for each action; humanoid_common simultaneously supports the SuperMimic action catalog (#31). The two commits are one minute apart, part of the same coordination.
The timeline is worth comparing: on November 5, humanoid_linglong’s #22 first brought LingLong’s SuperMimic standing policy into the configuration system—one robot model, one policy at the time. Six days later the same capability expanded to G1 and sank into the common-layer action catalog, showing it is moving from “policy configuration for a single model” to “a configurable feature across models”—once the action catalog is defined at the common layer, adding a new model only requires providing action implementations and bindings, without duplicating a scheduler.
For teams doing whole-machine integration, the significance of such work is turning “demos” into “product features”: standby, action selection, cancellation timing, audio prompts and resource preparation—all five together are what an action library should look like when a robot is delivered to a customer. Beijing’s three-year action plan issued the same day writes annual whole-machine capacity and mean time between failures into its assessment, and the configurability of the action library and the resettability of faults are part of production line takt time and after-sales cost.
Date: 2026-11-11 Source: spacemit-robotics/humanoid_unitree_g1, spacemit-robotics/humanoid_common
V. Summary and Trend Observations
1. Policy quantifies the demand side of embodied intelligence; RISC-V’s opportunity lies at both the joint and body levels. Beijing’s plan gives 200,000 units/year of whole-machine capacity, a 60% local supporting rate for key components, and dexterous hands under RMB 10,000—essentially writing “what chip, what cost, what reliability” into procurement parameters. The joint side needs hard real-time and deterministic networking; the body side needs compute and bandwidth to run VLA. Neither level is solved by a single large chip. Domestic RISC-V supply already has comparable products at both levels; the next competitive point is not peak compute, but whether cost can be pushed down to a level where whole-machine makers are willing to switch materials under the constraint of 2,000 hours MTBF.
2. The keyword for distributions this day is still “deliverability,” but the approach shifts from patches to matrices. RevyOS had 101 commits in one day, focused on consolidating board support for K3, EIC7700, SG2044, TH1520 and P550 into a defconfig matrix and hooking up both cross and native CI; openRuyi had 36 commits in one day, closing out both the observability and certification supply chains and upgrading the kernel to 7.3.4/6.18.63. Neither is doing new features; both are about “whether others can compile it and dare put it into production.”
3. Verification and observability become the new front line of toolchain competition. NEMU adds a fast path for AME matrix memory access, difftest vectorizes MMA verification, XiangShan adds an ISA smoke test to CI, and Espressif turns instruction-level trace into four pip-installable commands—four actions point to the same judgment: once cores and extensions become complex enough, the bottleneck shifts from “is the design right” to “can we quickly see what it is doing.” Teams underinvesting on this line will pay multiplied time in the integration phase.
4. XiangShan’s 18 pure refactoring commits show the high-performance core has entered a structural governance phase. Being able to do batch refactoring of a module that “changes only structure, not behavior” presupposes a dense enough regression surface and stable behavior. TLB and PTW are the hardest-to-debug part of the memory access path; consolidating permission checks and bypass results to a single exit is about leaving interfaces for future extensions. Such work does not appear in any performance number, but it determines the change cost when adding features to the next-generation core.
5. The capital-side pricing anchor is on the table, but asset boundaries remain to be disclosed. Alibaba announced a proposed spin-off of T-Head for a Hong Kong listing, and HKEX confirmed the split is permissible, with valuation ranges from $20–30 billion to $25–62 billion. For RISC-V companies, the reference value of this event is not the valuation figure but how the prospectus will separate external revenue from related-party transactions—the same standard will soon be used to measure all domestic compute chip companies, including RISC-V server and IP vendors still in the loss-making phase.
Appendix
Source Verification Table
| Source | Verification Result |
|---|---|
| Google News aggregation (24 Chinese and English query sets, 24-hour window) | 3 substantive new items within the window: the Beijing embodied AI robotics three-year action plan, pricing commentary on the Pingtouge spin-off announcement, and a Chinese reprint of MIPS Actus; the third was already covered on 11-09 and is not counted again. The ESP32-S31 running full Linux, Qwenputer, and Infineon AURIX RC1 items are all second-hand recirculation of old news, already covered in previous issues. Stock market quotes, fund NAVs, and sports betting items were removed wholesale |
| Hacker News / Phoronix / CNX Software / LinuxGizmos / RISC-V International | No new RISC-V-related items within the window; the most recent HN hit was the RISC-V computer project on Tang Nano 9K from 11-09, now out of date; Phoronix, CNX, and LinuxGizmos content for the period was unrelated to RISC-V; the riscv.org homepage currently promotes registration for the February 2027 San Francisco summit |
| GitHub organization and key repository checks (30+ organizations and repositories) | 12 organizations pushed within the window: espressif 42 repos, sophgo 19 repos, spacemit-com 8 repos, openRuyi-Project 8 repos, openhwfoundation 6 repos, sipeed 4 repos, spacemit-robotics 2 repos, Nuclei-Software 1 repo, revyos 1 repo; on the XiangShan side, three repos XiangShan / NEMU / difftest had commits, while XSCache and GEM5 had none; radxa-linkr, DAMO XuanTie (XUANTIE-RV), and ESWIN (eswincomputing) had zero pushes |
| riscv100.top inspection | No changes: the homepage What’s New, Upcoming, committee members, and the six SPEC CPU2006/2017/2026 leaderboards under the cross-architecture and RISC-V channels all showed no new records or score changes, so no space is allocated in the main text |
| RISC-V member website inspection (391 sites) | 323 successfully crawled, 8 reported content changes: MIT, KTH, University of Athens, LNCC, PES, and the openEuler community were homepage rotations or announcement updates; PerfXLab was a connection timeout page and QuantumShift a security verification page; none had substantive product developments and none were included; 68 failed to crawl (including ISCAS, Tencent, Ingenic, MediaTek, DENSO); failures do not block |
| openRuyi official news page and main repo | 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 main repo’s 36 commits have been included per the log, with the kernel-team-tools tag progression merged into the same item |
Complete Source List
- [1] Wallstreetcn: Beijing embodied AI three-year plan released, 200,000 units annual capacity and 50 benchmark scenarios — https://awtmt.com/articles/3785725
- [2] Beijing Business Today: Beijing robotics industry plan unveiled (Sohu reprint) — https://m.sohu.com/a/1100274218_115865
- [3] revyos/linux repository — https://github.com/revyos/linux
- [4] openRuyi-Project/openRuyi main repo — https://github.com/openRuyi-Project/openRuyi
- [5] openRuyi-Project/kernel-team-tools — https://github.com/openRuyi-Project/kernel-team-tools
- [6] espressif/esp-riscv-trace — https://github.com/espressif/esp-riscv-trace
- [7] Adafruit: A RISC-V-only clang + lld for the CH32 Arduino core — https://blog.adafruit.com/2026/11/11/a-risc-v-only-clang-lld-for-the-ch32-arduino-core/
- [8] ch32-riscv-ug/ch32rv-clang — https://github.com/ch32-riscv-ug/ch32rv-clang
- [9] OpenXiangShan/NEMU — https://github.com/OpenXiangShan/NEMU
- [10] OpenXiangShan/difftest — https://github.com/OpenXiangShan/difftest
- [11] openRuyi-Project/ruyi-deck — https://github.com/openRuyi-Project/ruyi-deck
- [12] 36Kr: After burning 190 billion, Alibaba pushes Pingtouge to market — https://m.36kr.com/p/4065139957192322
- [13] Eastmoney reprint of Time Weekly: Pingtouge sprints independently toward a Hong Kong IPO — https://finance.eastmoney.com/a/202611113928531365.html
- [14] OpenXiangShan/XiangShan repository — https://github.com/OpenXiangShan/XiangShan
- [15] spacemit-com/llama.cpp — https://github.com/spacemit-com/llama.cpp
- [16] spacemit-com/ai-sdk — https://github.com/spacemit-com/ai-sdk
- [17] spacemit-com/model-zoo-vision — https://github.com/spacemit-com/model-zoo-vision
- [18] sophgo/cviruntime — https://github.com/sophgo/cviruntime
- [19] Nuclei-Software/nuclei-tool-guide — https://github.com/Nuclei-Software/nuclei-tool-guide
- [20] sipeed/SLogicWeb — https://github.com/sipeed/SLogicWeb
- [21] sipeed/TangMega-138K-example — https://github.com/sipeed/TangMega-138K-example
- [22] spacemit-robotics/humanoid_unitree_g1 — https://github.com/spacemit-robotics/humanoid_unitree_g1
- [23] spacemit-robotics/humanoid_common — https://github.com/spacemit-robotics/humanoid_common
Limitations: This issue covers a 24-hour window from 05:00 on November 11 to 05:00 on November 12. The effective new items on the news side are concentrated in two: the Beijing embodied AI three-year action plan and pricing commentary on the Pingtouge spin-off announcement; the rest are second-hand recirculation of old news, removed per the deduplication rules; the Chinese reprint of MIPS Actus appeared on 11-11, but the platform itself was already included as an item on 11-09, so it is not repeated in this issue. On the GitHub side, the focus is mainly distributions and board support: revyos/linux 101 commits, openRuyi 36 commits, XiangShan 21 commits, plus batch pushes from 19 sophgo repos and 42 espressif repos; the latter two are release trains and upstream syncs and were not broken out item by item. riscv100.top showed no changes this issue, so no archive was generated. Of the 391 member websites inspected, 68 failed to crawl (including ISCAS, Tencent, Ingenic, MediaTek, DENSO, etc.); of the 8 judged to have content changes, 2 were actually timeout and security verification pages. On the Chinese side, queries such as “XuanTie,” “Sophgo,” “Nuclei,” and “ESWIN” still returned mostly stock market articles, which were removed based on headline semantics.