--- license: apache-2.0 base_model: FireRedTeam/FireRedASR2-LLM pipeline_tag: automatic-speech-recognition library_name: openasr tags: - automatic-speech-recognition - speech-to-text - openasr - oasr - firered2-llm ---
# FireRedASR2 LLM Β· OpenASR **FireRedTeam's LLM-backbone Mandarin-first bilingual ASR β€” 8B+ parameters engineered for state-of-the-art Chinese and dialect accuracy** [![License](https://img.shields.io/badge/license-Apache--2.0-2563eb.svg)](https://huggingface.co/FireRedTeam/FireRedASR2-LLM) [![Format](https://img.shields.io/badge/format-.oasr-7c3aed.svg)](https://github.com/QuintinShaw/openasr) [![Runtime](https://img.shields.io/badge/runtime-OpenASR-111827.svg)](https://openasr.org) [![Base model](https://img.shields.io/badge/base-FireRedASR2--LLM-f59e0b.svg)](https://huggingface.co/FireRedTeam/FireRedASR2-LLM) Native speech-to-text in the **[OpenASR](https://github.com/QuintinShaw/openasr)** runtime β€” engineered for peak performance on CPU & GPU, **no Python at inference time**.
--- ## ✨ Highlights - πŸ₯‡ **Best-in-class Mandarin accuracy** β€” 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) - πŸ—£οΈ **Leading dialect and accent coverage** β€” 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) - πŸ‡¨πŸ‡³πŸ‡¬πŸ‡§ **Bilingual Mandarin + English** β€” one 8B+ parameter checkpoint handles both languages, as demonstrated in the upstream model card's bilingual examples (FireRedTeam's FireRedASR2-LLM model card) - 🧠 **LLM-scale decoder backbone** β€” 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) - πŸ”’ **Fully local, q4_k build** β€” 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) - πŸ¦€ **Native in OpenASR** β€” `.oasr` packs run with no Python at inference, engineered for peak performance on CPU & GPU ## πŸš€ Quickstart ```bash # 1. Install the OpenASR CLI Β· https://openasr.org # 2. Pull a build (pick a quant β€” see the table below) openasr pull firered2-llm:q4 # 3. Transcribe openasr transcribe audio.wav --model firered2-llm ``` All builds for this model: ```bash openasr pull firered2-llm:q4 ``` ## πŸ“¦ Available builds | Quant | File (`.oasr`) | Size | RAM peak | RTF Β· M1 CPU | RTF Β· M1 GPU | JFK Ξ”WER vs fp16 | |:------|:---------------|-----:|---------:|-------------:|-------------:|-----------------:| | q4_k | `firered2-llm-q4_k.oasr` | 5.10 GB | 9.39 GB | 0.77Γ— | 0.40Γ— | n/a | RTF = real-time factor on the fixed 11s JFK clip (**lower is faster**); RAM peak measured per pack in an isolated subprocess. JFK Ξ”WER compares each quantized build's JFK transcript to this model's fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy. **q4_k** is the recommended default β€” near-reference quality at a fraction of the footprint. ## 🧠 About FireRedASR2 LLM FireRedASR2-LLM is the Encoder-Adapter-LLM member of **FireRedASR2**, the successor to FireRedTeam's open-source industrial-grade **FireRedASR** speech-recognition family, released as part of the **FireRedASR2S** all-in-one ASR system. At 8B+ parameters with an LLM-scale decoder backbone, it is the bigger and more accurate sibling of the already-available **firered-aed-l-v2** (FireRedASR2-AED, a 1.1B-parameter attention encoder-decoder reporting 3.05%/11.67% avg CER). The FireRedASR2S technical report (arXiv:2603.10420) and the upstream model card both report **2.89% average Character Error Rate** across four public Mandarin benchmarks and **11.55% average CER** across 19 public Chinese dialect/accent benchmarks -- outperforming Doubao-ASR (3.69%/15.39%), Qwen3-ASR (3.76%/11.85%), and Fun-ASR (4.16%/12.76%) on the paper's own comparison table. These numbers belong to FireRedASR2-LLM and should not be confused with the smaller firered-aed-l-v2's 3.05%/11.67% figures. The model is bilingual (Mandarin Chinese and English). This OpenASR pack distributes the q4_k quantization build, repackaged as an `.oasr` pack that runs natively in the OpenASR runtime -- fully local, no Python at inference time, no cloud upload. Licensed under Apache-2.0, inherited from the upstream release. ## βš™οΈ How these packs were made Converted from [FireRedTeam/FireRedASR2-LLM](https://huggingface.co/FireRedTeam/FireRedASR2-LLM) with the OpenASR importer: ```bash openasr model-pack import firered-llm .oasr \ --package-id firered2-llm --quantization {fp16,q8-0,q4-k} ``` The `.oasr` container is GGUF-backed; packs use zero-copy mmap weight binding and graph buffer reuse to keep peak memory low. ## βš–οΈ License These packs **inherit the upstream model's license: Apache-2.0** ([source](https://huggingface.co/FireRedTeam/FireRedASR2-LLM)). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization. ## πŸ™ Acknowledgements This pack is a redistribution of **FireRedASR2-LLM**, created and released by **FireRedTeam** ([FireRedTeam/FireRedASR2-LLM](https://huggingface.co/FireRedTeam/FireRedASR2-LLM), [FireRedTeam/FireRedASR2S](https://github.com/FireRedTeam/FireRedASR2S)). All credit for the architecture, training, and weights belongs to FireRedTeam; the license is inherited from and identical to the upstream model (**Apache-2.0**, as declared on the upstream model card). Thank you to FireRedTeam for releasing their work openly. OpenASR only performs format conversion, quantization, runtime verification, and local-inference adaptation. ## πŸ”— Links - πŸ¦€ **OpenASR** β€” - 🌐 **Website** β€” - πŸ€— **Upstream model** β€” [FireRedTeam/FireRedASR2-LLM](https://huggingface.co/FireRedTeam/FireRedASR2-LLM)