OpenASR/firered2-llm
04.7k
1---2license: apache-2.03base_model: FireRedTeam/FireRedASR2-LLM4pipeline_tag: automatic-speech-recognition5library_name: openasr6tags:7 - automatic-speech-recognition8 - speech-to-text9 - openasr10 - oasr11 - firered2-llm12---13 14<div align="center">15 16# FireRedASR2 LLM · OpenASR17 18**FireRedTeam's LLM-backbone Mandarin-first bilingual ASR — 8B+ parameters engineered for state-of-the-art Chinese and dialect accuracy**19 20[](https://huggingface.co/FireRedTeam/FireRedASR2-LLM)21[](https://github.com/QuintinShaw/openasr)22[](https://openasr.org)23[](https://huggingface.co/FireRedTeam/FireRedASR2-LLM)24 25Native speech-to-text in the **[OpenASR](https://github.com/QuintinShaw/openasr)** runtime —26engineered for peak performance on CPU & GPU, **no Python at inference time**.27 28</div>29 30---31 32## ✨ Highlights33 34- 🥇 **Best-in-class Mandarin accuracy** — 2.89% average CER across four public Mandarin benchmarks, outperforming Doubao-ASR (3.69%), Qwen3-ASR (3.76%), and Fun-ASR (4.16%) on the same comparison table (arXiv:2603.10420; FireRedTeam's FireRedASR2-LLM model card)35- 🗣️ **Leading dialect and accent coverage** — 11.55% average CER across 19 public Chinese dialect/accent benchmarks, ahead of Doubao-ASR (15.39%) and Qwen3-ASR (11.85%) (arXiv:2603.10420; FireRedTeam's FireRedASR2-LLM model card)36- 🇨🇳🇬🇧 **Bilingual Mandarin + English** — one 8B+ parameter checkpoint handles both languages, as demonstrated in the upstream model card's bilingual examples (FireRedTeam's FireRedASR2-LLM model card)37- 🧠 **LLM-scale decoder backbone** — Encoder-Adapter-LLM architecture at 8B+ parameters, the bigger and more accurate sibling of the already-available firered-aed-l-v2 (1.1B AED, 3.05%/11.67% avg CER) (FireRedTeam's FireRedASR2-LLM model card)38- 🔒 **Fully local, q4_k build** — runs 100% on-device via OpenASR's .oasr runtime with no cloud upload, under the Apache-2.0 license inherited from upstream (FireRedTeam's FireRedASR2-LLM model card)39- 🦀 **Native in OpenASR** — `.oasr` packs run with no Python at inference, engineered for peak performance on CPU & GPU40 41## 🚀 Quickstart42 43```bash44# 1. Install the OpenASR CLI · https://openasr.org45# 2. Pull a build (pick a quant — see the table below)46openasr pull firered2-llm:q447 48# 3. Transcribe49openasr transcribe audio.wav --model firered2-llm50```51 52All builds for this model:53 54```bash55openasr pull firered2-llm:q456```57 58## 📦 Available builds59 60| Quant | File (`.oasr`) | Size | RAM peak | RTF · M1 CPU | RTF · M1 GPU | JFK ΔWER vs fp16 |61|:------|:---------------|-----:|---------:|-------------:|-------------:|-----------------:|62| q4_k | `firered2-llm-q4_k.oasr` | 5.10 GB | 9.39 GB | 0.77× | 0.40× | n/a |63 64<sub>RTF = real-time factor on the fixed 11s JFK clip (**lower is faster**); RAM peak measured per pack65in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's66fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy.67**q4_k** is the recommended default — near-reference quality at a fraction of the68footprint.</sub>69 70## 🧠 About FireRedASR2 LLM71 72FireRedASR2-LLM is the Encoder-Adapter-LLM member of **FireRedASR2**, the successor to73FireRedTeam's open-source industrial-grade **FireRedASR** speech-recognition family, released as74part of the **FireRedASR2S** all-in-one ASR system. At 8B+ parameters with an LLM-scale decoder75backbone, it is the bigger and more accurate sibling of the already-available76**firered-aed-l-v2** (FireRedASR2-AED, a 1.1B-parameter attention encoder-decoder reporting773.05%/11.67% avg CER). The FireRedASR2S technical report (arXiv:2603.10420) and the upstream model78card both report **2.89% average Character Error Rate** across four public Mandarin benchmarks79and **11.55% average CER** across 19 public Chinese dialect/accent benchmarks -- outperforming80Doubao-ASR (3.69%/15.39%), Qwen3-ASR (3.76%/11.85%), and Fun-ASR (4.16%/12.76%) on the paper's own81comparison table. These numbers belong to FireRedASR2-LLM and should not be confused with the82smaller firered-aed-l-v2's 3.05%/11.67% figures. The model is bilingual (Mandarin Chinese and83English). This OpenASR pack distributes the q4_k quantization build, repackaged as an `.oasr`84pack that runs natively in the OpenASR runtime -- fully local, no Python at inference time, no85cloud upload. Licensed under Apache-2.0, inherited from the upstream release.86 87## ⚙️ How these packs were made88 89Converted from [FireRedTeam/FireRedASR2-LLM](https://huggingface.co/FireRedTeam/FireRedASR2-LLM) with the OpenASR importer:90 91```bash92openasr model-pack import firered-llm <src> <out>.oasr \93 --package-id firered2-llm --quantization {fp16,q8-0,q4-k}94```95 96The `.oasr` container is GGUF-backed; packs use zero-copy mmap weight binding and graph97buffer reuse to keep peak memory low.98 99## ⚖️ License100 101These packs **inherit the upstream model's license: Apache-2.0**102([source](https://huggingface.co/FireRedTeam/FireRedASR2-LLM)). OpenASR packaging retains the upstream copyright and103NOTICE; the only modifications are format conversion and quantization.104 105## 🙏 Acknowledgements106 107This pack is a redistribution of **FireRedASR2-LLM**, created and released by **FireRedTeam**108([FireRedTeam/FireRedASR2-LLM](https://huggingface.co/FireRedTeam/FireRedASR2-LLM),109[FireRedTeam/FireRedASR2S](https://github.com/FireRedTeam/FireRedASR2S)). All credit for the110architecture, training, and weights belongs to FireRedTeam; the license is inherited from and111identical to the upstream model (**Apache-2.0**, as declared on the upstream model card). Thank112you to FireRedTeam for releasing their work openly. OpenASR only performs format conversion,113quantization, runtime verification, and local-inference adaptation.114 115## 🔗 Links116 117- 🦀 **OpenASR** — <https://github.com/QuintinShaw/openasr>118- 🌐 **Website** — <https://openasr.org>119- 🤗 **Upstream model** — [FireRedTeam/FireRedASR2-LLM](https://huggingface.co/FireRedTeam/FireRedASR2-LLM)120