FunASR llama.cpp runtime v0.2.0

Prebuilt, self-contained FunASR llama.cpp / GGUF binaries for Fun-ASR-Nano, SenseVoiceSmall, and Paraformer, with built-in FSMN-VAD. No Python ASR runtime or local C++ build is required.

Highlights

• SRT subtitle output is available across Fun-ASR-Nano, SenseVoiceSmall, and Paraformer. Pass --srt to emit timestamped subtitle cues; the default plain-text output remains unchanged.
• The bundled llama.cpp revision refreshes Vulkan submission batching for smaller AMD GPUs and improves VK_ERROR_DEVICE_LOST diagnostics.
• AMD Vulkan troubleshooting now documents GGML_VK_MAX_NODES_PER_SUBMIT and GGML_VK_SERIALIZE_SUBMISSIONS, plus a CPU fallback. These changes improve diagnostics and batching behavior, but this release does not claim that every vendor-driver crash is fixed; please report the GPU model, driver version, command, and full log if a device still fails.

Download

Nine platform archives are attached:

• Linux: linux-arm64, linux-x64, linux-x64-avx2, linux-x64-vulkan
• macOS: macos-arm64
• Windows: windows-x64, windows-x64-avx2, windows-x64-vulkan, windows-x64-cuda

Use the default x64 archive for maximum CPU compatibility. Choose x64-avx2 on CPUs with AVX2, FMA, F16C, and BMI2. The Vulkan archives require a working Vulkan driver/ICD and currently accelerate the SenseVoiceSmall graph with llama-funasr-sensevoice ... --backend vulkan. The Windows CUDA archive targets CUDA architecture 86 and requires a compatible NVIDIA driver; build from source for other GPU architectures.

Quick Start

Download the default quantized model with:

bash download-funasr-model.sh <sensevoice|paraformer|nano>

The helper requires the Hugging Face CLI (pip install -U huggingface_hub). Then run llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

Full documentation: https://github.com/modelscope/FunASR/blob/runtime-llamacpp-v0.2.0/runtime/llama.cpp/README.md

FunASR 1.4.1

Install or upgrade the Python package:

pip install -U funasr==1.4.1

Python package

• Routes the Hugging Face paraformer-en alias to the official English checkpoint instead of silently downloading the Chinese model.
• Adds LoRA fine-tuning support for the Qwen3 LLM inside Fun-ASR-Nano, including deterministic regression coverage and bilingual guidance.
• Avoids checkpoint-selection failures when a save step has not yet produced validation metrics.

Tagged source

• Adds stable JSONL transcript and timestamp output to the native ONNX Runtime offline and two-pass binaries.
• Adds a reproducible SenseVoice TensorRT/Triton deployment path.
• Adds the openclaw-funasr realtime transcription provider for OpenClaw Talk and Voice Call.
• Includes native runtime buffer-ownership fixes merged after v1.4.0.

Prebuilt llama.cpp runtime

The release includes the same verified FunASR llama.cpp / GGUF runtime builds for nine Linux, macOS, and Windows targets. Use SHA256SUMS-v1.4.1 to verify every Python and runtime artifact after download.

Full Changelog: v1.4.0...v1.4.1

FunASR 1.4.0

Install or upgrade the Python package:

pip install -U funasr==1.4.0

Python package

• Rejects the common vda_model misspelling immediately and points users to the supported vad_model argument, instead of silently running without VAD.
• Restores SenseVoice English spelling normalization in installed wheels by packaging english.json.
• Restores RWKV-BAT CUDA kernel compilation from installed wheels by packaging the encoder and decoder .cpp and .cu sources.

Tagged source

• Finalizes legacy WebSocket file-input sessions correctly when the client sends the end-control message. This runtime is distributed in the tagged GitHub source and is not part of the PyPI wheel.

Prebuilt llama.cpp runtime

The release assets include self-contained FunASR llama.cpp / GGUF runtimes for SenseVoice, Paraformer, and Fun-ASR-Nano:

• Linux arm64
• Linux x64
• Linux x64 AVX2
• Linux x64 Vulkan
• macOS arm64
• Windows x64
• Windows x64 AVX2
• Windows x64 CUDA
• Windows x64 Vulkan

Use SHA256SUMS-v1.4.0 to verify every Python and runtime artifact after download.

Full Changelog: v1.3.30...v1.4.0

What's Changed

• ci: guard FunClip v2.1.0 launch pages by @LauraGPT in #3417
• Add MCP Toplist rank badge by @chrstphe in #3418
• fix: limit vLLM VAD segments to 30 seconds by @LauraGPT in #3420
• test: guard donor navigation across public pages by @LauraGPT in #3421
• test: harden donor navigation website contract by @LauraGPT in #3422
• Improve MCP discovery and release 0.1.2 by @LauraGPT in #3423
• feat(site): launch the FunASR deployment product hub by @LauraGPT in #3424
• docs: surface the deployment product hub by @LauraGPT in #3425
• fix(server): allow trusted browser origins by @LauraGPT in #3427
• feat(site): attribute repository conversions by @LauraGPT in #3428
• docs(site): surface Orca SenseVoice desktop integration by @LauraGPT in #3431
• docs: add RTX 4090 Fun-ASR community benchmark by @LauraGPT in #3432
• fix(server): return speaker labels for spk requests by @LauraGPT in #3433
• fix(auto-model): preserve VAD segments on punctuation mismatch by @LauraGPT in #3434
• Decode container-formatted audio bytes before inference by @LauraGPT in #3435
• chore: release FunASR 1.3.30 by @LauraGPT in #3436

New Contributors

• @chrstphe made their first contribution in #3418

Full Changelog: v1.3.29...v1.3.30

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.1.9.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.3.30"

FunASR MCP Server 0.1.2 improves agent discovery and updates the packaged runtime.

Highlights

• Gives MCP clients explicit guidance on when to use or avoid the transcribe_audio tool.
• Documents local/container path scope, supported language hints, first-call model download and cache behavior, and transcript segment output.
• Updates the container runtime from FunASR 1.3.14 to FunASR 1.3.29.
• Publishes a multi-platform public image and refreshed official MCP Registry metadata.

Install

docker pull ghcr.io/modelscope/funasr-mcp:0.1.2

Mount audio read-only at /audio and persist the ModelScope cache as documented in examples/mcp_server/README.md.

• Official MCP Registry: https://registry.modelcontextprotocol.io/v0.1/servers?search=io.github.modelscope/funasr-mcp
• GHCR package: https://github.com/orgs/modelscope/packages/container/package/funasr-mcp
• Documentation: https://github.com/modelscope/FunASR/tree/main/examples/mcp_server
• Changes: #3423

What's Changed

• test: guard v1.3.28 website launch by @LauraGPT in #3410
• docs: surface official Dify Marketplace plugin by @LauraGPT in #3411
• docs: use QwenAudio canonical repository links by @LauraGPT in #3412
• Return VAD sentence timestamps for SenseVoice by @LauraGPT in #3414
• chore: release FunASR 1.3.29 by @LauraGPT in #3415

Full Changelog: v1.3.28...v1.3.29

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.1.9.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.3.29"

Highlights

• Realtime WebSocket finalization now preserves the best continuous transcript when a VAD-locked decode truncates, repeats a hallucinated phrase, or raises.
• STOP, VAD finalization, and speaker completion share the same reliable path; WebSocket disconnect and error handling no longer depends on lazy module attributes.
• SenseVoice subtitle segmentation now aligns rich tags, punctuation, and word/BPE timestamps without collapsing Chinese into one cue or damaging English text.

Install

python -m pip install -U "funasr==1.3.28"

PyPI: https://pypi.org/project/funasr/1.3.28/

Verification

• 118 focused regression tests passed from the exact merge commit.
• Wheel and source distribution passed Twine checks and isolated-install version checks.
• The release tag is SSH-signed and points to merge commit 689cc94.

Full Changelog: v1.3.27...v1.3.28

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.1.9.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.3.28"

Weekly Python release focused on OpenAI-compatible serving reliability.

Highlights

• verbose_json now reports the language detected by SenseVoice instead of silently defaulting to Chinese.
• Fun-ASR-Nano reuses its cached AutoModel after a vLLM fallback, avoiding repeated fallback construction on every request.
• Failed vLLM/VAD setup no longer leaves half-initialized engine state when both setup and fallback fail, so a later request can retry.
• The OpenAPI metadata and funasr-server startup banner now follow the installed package version.

Install

python -m pip install -U "funasr==1.3.27"

PyPI: https://pypi.org/project/funasr/1.3.27/

The wheel and source distribution are also attached to this GitHub release. Prebuilt llama.cpp / GGUF downloads for Linux, macOS, and Windows are listed below.

Validation

• 41 server, CLI, documentation, packaging, and runtime contract tests passed.
• Wheel and sdist passed twine check.
• A clean public PyPI install reports package, CLI, and server version 1.3.27.

Full changelog: v1.3.26...v1.3.27

Runtime downloads

This Python release pairs with the current prebuilt llama.cpp / GGUF runtime release: runtime-llamacpp-v0.1.9.

The same verified runtime assets are attached directly to this Python release so users can find the package and self-contained llama-funasr-* binaries in one place.

Quick start: download one asset, unpack it, then run the bundled download-funasr-model.sh <sensevoice|paraformer|nano> helper and one of llama-funasr-cli, llama-funasr-sensevoice, or llama-funasr-paraformer.

For Python users, install from PyPI:

python -m pip install -U "funasr==1.3.27"

Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime: SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD. Download the default quantized model with bash download-funasr-model.sh <sensevoice|paraformer|nano> (the helper requires the Hugging Face CLI: pip install -U huggingface_hub), then run llama-funasr-cli / llama-funasr-sensevoice / llama-funasr-paraformer. Use the default x64 asset for maximum CPU compatibility; use the x64-avx2 asset on CPUs with AVX2/FMA/F16C/BMI2 for higher throughput. The Vulkan assets are linux-x64-vulkan and windows-x64-vulkan; they require a working Vulkan driver/ICD and enable SenseVoiceSmall graph execution with llama-funasr-sensevoice ... --backend vulkan. Build from source with -DGGML_VULKAN=ON to validate platform-specific GPU stacks. The Windows CUDA asset is windows-x64-cuda; it requires an NVIDIA driver compatible with the CUDA Toolkit version configured by the release workflow, targets CUDA architecture 86, and enables SenseVoiceSmall graph execution with llama-funasr-sensevoice ... --backend cuda. Build from source for other GPU architectures. No Python ASR runtime or local build is required. Docs: https://github.com/modelscope/FunASR/blob/runtime-llamacpp-v0.1.9/runtime/llama.cpp/README.md

FunASR MCP server v0.1.1 is the first fully published, installable release of the official FunASR Model Context Protocol server. It exposes local FunASR/SenseVoice transcription as the transcribe_audio MCP tool over stdio.

Install

The public OCI image is available for Linux amd64:

docker pull ghcr.io/modelscope/funasr-mcp:0.1.1

MCP clients can launch it with a read-only audio directory and a persistent model cache:

docker run --rm -i       --mount type=bind,src=/path/to/audio,dst=/audio,readonly       --mount type=volume,src=funasr-mcp-cache,dst=/root/.cache/modelscope       ghcr.io/modelscope/funasr-mcp:0.1.1

Pass files to the tool as /audio/. The default server uses iic/SenseVoiceSmall; model downloads stay in the named Docker volume between runs.

Registry

• Official MCP Registry name: io.github.modelscope/funasr-mcp
• OCI package: ghcr.io/modelscope/funasr-mcp:0.1.1
• Source and configuration: https://github.com/modelscope/FunASR/tree/mcp-v0.1.1/examples/mcp_server
• Container package page: https://github.com/orgs/modelscope/packages/container/package/funasr-mcp

The release pipeline passed metadata/schema validation, container build, MCP initialize/tools-list smoke tests, anonymous GHCR pull verification, GitHub OIDC publication, and official Registry listing verification.

Related downloads

• Python/PyPI release: https://github.com/modelscope/FunASR/releases/tag/v1.3.26
• Prebuilt llama.cpp/GGUF binaries for Linux, macOS, and Windows: https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.8
• MCP server documentation: https://github.com/modelscope/FunASR/blob/mcp-v0.1.1/examples/mcp_server/README.md

Changes

• Publish versioned and latest OCI images to GHCR.
• Publish canonical server metadata to the official MCP Registry.
• Verify public pull access before Registry publication.
• Keep organization package-visibility API limitations from blocking an already-public image.
• Preserve a read-only host audio mount and persistent ModelScope model cache.