RISC-V Open-Source Maintainer Stance & Receptivity Analysis

Date: August 2026
Scope: Evaluation of maintainer support, PR receptivity, and ecosystem friction across 147 open-source projects (reports directory).

Executive Summary

Across the 147 open-source projects tracked in this repository, maintainer attitudes toward RISC-V (riscv64) fall into three distinct categories:

  1. Champions (Gold Standard): Projects where upstream maintainers actively embrace RISC-V, maintain dedicated machine maintainers, gate PRs with native/official CI, and publish binary releases.
  2. Progressing (Traction with Operational Friction): Projects with merged RISC-V code and active community momentum, but where maintainers leave testing to community runners or lag in publishing prebuilt PyPI wheels.
  3. Resistant / Hostile / Neglected: Projects where maintainers have closed or ignored RISC-V PRs, explicitly reverted prebuilt binary wheels, placed issues in “Cold Storage”, or refused to assign maintainer review capacity.

1. Champion Projects (Working the Best with RISC-V)

These projects feature Tier-1/2 upstream maintainer backing, dedicated machine maintainers, active native or QEMU CI gating, prompt PR reviews, and zero maintainer friction:

  • Linux Kernel & GNU Toolchain (glibc, GCC, GDB, linux-perf, eBPF):
    • glibc has three active named machine maintainers (Palmer Dabbelt / Rivos, Andrew Waterman / SiFive, Peter Bergner / IBM). Full mainline feature parity.
  • LLVM Compiler Infrastructure & Clang (LLDB):
    • riscv64 is a first-class compilation target with complete RVV 1.0 vector intrinsic and auto-vectorization support.
  • Go / Golang (go):
    • Official secondary port with active community builders, clean toolchain support (GOARCH=riscv64), and 1:1 runtime parity for web/cloud microservices.
  • OpenJDK / Java (openjdk):
    • JEP 422 (“Linux/RISC-V Port”) delivered mainline into OpenJDK 19+. HotSpot JIT compiler is fully supported and maintained.
  • llama.cpp (llama-cpp):
    • Community-driven C++ LLM inference engine. Rapidly merged RVV 1.0 matrix multiplication and dequantization kernels (VLEN=128/256) with zero maintainer friction.
  • OpenSSL (openssl):
    • Full upstream integration including RVV-accelerated cryptographic primitives (Zkn/Zks).
  • Container Infrastructure (containerd, runc, Docker, Kubernetes):
    • High maintainer receptivity with full multi-arch container image support (linux/riscv64).
  • FFmpeg (ffmpeg) & dav1d (dav1d):
    • Active RVV assembly vectorization PRs merged regularly by core video codec maintainers.

2. Progressing Projects (Traction with Operational Friction)

These projects have working code and merged upstream PRs, but suffer from operational friction such as missing PyPI binary wheels, opt-in/cross-compiled CI, or unmerged vectorization routines:

  • vLLM (vllm):
    • In-tree CPU backend merged with RVV attention kernels (VLEN=128 and VLEN=256), reviewed/approved by Intel maintainer bigPYJ1151. However, it lacks automated upstream CI and prebuilt PyPI wheels.
  • NumPy (numpy):
    • Merged native RISC-V CI (PR #31488, May 2026); an official riscv64 PyPI wheel is targeted for release 2.6.0 (Q3 2026).
  • LangChain (langchain):
    • Pure-Python core (py3-none-any wheel) installs cleanly, but is blocked by missing binary wheels in downstream C++/Rust dependencies.
  • Chromium / V8 & Firefox / SpiderMonkey:
    • Functional RVV JIT backends and desktop browser rendering, but upstream Google and Mozilla teams treat RISC-V as community-maintained without mandatory CI gating.
  • oneDNN (onednn) & OpenBLAS (openblas):
    • oneDNN v3.10+ merged RISC-V CPU backend support (~8.85x speedup for elementwise ops). OpenBLAS has active RVV kernel work, but suffers from compiler edge bugs and unmerged LAPACK routines.
  • SQLAlchemy & tokenizers:
    • SQLAlchemy merged QEMU CI; tokenizers publishes manylinux_2_31_riscv64 wheels (though excluding mimalloc).

3. Hostile, Resistant, or Neglected Projects

These projects exhibit active maintainer resistance, abandoned PRs, closed issues without merge, reverted binary wheels, or structural neglect:

  1. PyTorch (pytorch/pytorch) (pytorch):
    • Maintainer Stance: Meta maintainers (primarily malfet) treat RISC-V as an opt-in, non-blocking community effort.
    • Friction:
      • Placed the primary RISC-V CI request in “Cold Storage” (Issue #141550).
      • Force-merges RISC-V PRs without maintainer review (“lint is green, rest is not compiled”).
      • Refuses to assign a CODEOWNERS entry for RISC-V.
      • Leaves the core ATen RVV vectorization (PR #175746) stalled for months over scalable-vector copy debates (Vectorized::size()).
      • Zero riscv64 PyPI wheels.
  2. google/sentencepiece (vllm report):
    • Maintainer Stance: Maintainer explicitly reverted a working riscv64 PyPI wheel build (PR #1226), leaving Issue #1250 open. This directly blocks standard pip install for LLaMA and Gemma tokenizers on RISC-V.
  3. openai/tiktoken (vllm report):
    • Maintainer Stance: OpenAI maintainers have ignored and stalled the riscv64 CI PR (#506) since March 2026 with no response, blocking GPT-2/GPT-4 tokenization wheels.
  4. numba / llvmlite (numba):
    • Maintainer Stance: Maintainers have ignored and closed riscv64 support requests (llvmlite #923). JIT Python numeric acceleration remains completely unsupported on RISC-V.
  5. pydantic / pydantic-core (langchain report):
    • Maintainer Stance: Closed the riscv64 CI PR (#1901) without merging.
  6. protocolbuffers / protobuf (pytorch report):
    • Maintainer Stance: Abandoned prebuilt protoc compiler binary PRs for riscv64 (#23206 / #23205), forcing users to build protoc from source.
  7. XNNPACK (google/XNNPACK) (xnnpack):
    • Maintainer Stance: Unconditionally enabled the Zvfh (FP16) compiler flag (PR #9516) without runtime hardware detection via cpuinfo, causing 100+ FP16 test failures (Issue #9886). Excluded riscv64 from default CMake dependencies without explicit user flags.
  8. FBGEMM / NNPACK / psimd:
    • Maintainer Stance: Meta/upstream maintainers explicitly restrict FBGEMM to x86/AArch64, while NNPACK and psimd are archived or unmaintained.

Maintainer Stance Summary Matrix

Project Maintainer Stance Upstream Status Key Friction Point / Hostility Signal
Linux / glibc / GCC 🟢 Champions Tier 1 / Mainline Full maintainer bench; zero friction.
Go / Rust / OpenJDK 🟢 Champions Tier 1/2 Mainline Active builders; JEP 422 delivered.
llama.cpp 🟢 Champions Mainline Merged Fast reviews; RVV 1.0 kernels merged cleanly.
NumPy 🟡 Traction Native CI Merged Wheel targeted Q3 2026 (release 2.6.0).
vLLM 🟡 Traction In-Tree CPU Backend Merged RVV attention; lacks automated CI & wheels.
PyTorch 🔴 Hostile / Resistant Tier 3 (Opt-in) Issues placed in “Cold Storage”; force-merged PRs; unmerged ATen RVV.
sentencepiece 🔴 Hostile / Resistant Wheel Reverted Explicitly reverted riscv64 PyPI wheel PR #1226.
tiktoken 🔴 Hostile / Neglected Stalled PR OpenAI ignored CI PR #506 since March 2026.
numba / llvmlite 🔴 Hostile / Neglected Issues Closed Closed riscv64 support request #923 without action.
protobuf 🔴 Hostile / Neglected Abandoned Abandoned prebuilt protoc binary PR #23206.