--- title: "Liquid AI's LFM2.5 Beats 7.7B Japanese Voice Models at 1.5B Parameters" subtitle: "Liquid AI releases two sub-2B Japanese models: a first-of-its-kind end-to-end speech model and an updated language model that tops the sub-2B Japanese benchmark charts." url: "https://alphasignal.ai/news/liquid-ai-s-lfm2-5-beats-7-7b-japanese-voice-models-at-1-5b-parameters" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Liquid AI" company_url: "https://alphasignal.ai/company/liquidai" published: "2026-06-06T03:07:57.000Z" updated: "2026-06-16T20:15:00.488Z" type: "News" topics: - "Audio" - "Llms" subtopics: - "Text To Speech" - "Speech To Text" - "Small Models" upvotes: 1841 read_time: "5 min" source_url: "https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP-202606" comments: 0 --- # Liquid AI's LFM2.5 Beats 7.7B Japanese Voice Models at 1.5B Parameters > Liquid AI releases two sub-2B Japanese models: a first-of-its-kind end-to-end speech model and an updated language model that tops the sub-2B Japanese benchmark charts. Liquid AI just dropped two new Japanese-language models, and the headline number is hard to ignore: a 1.5B-parameter audio model that outperforms a 7.7B-parameter competitor on conversational Japanese benchmarks. The release covers two distinct models , [LFM2.5-Audio-1.5B-JP](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-JP), an end-to-end speech model, and [LFM2.5-1.2B-JP-202606](https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP-202606), an updated general-purpose Japanese chat model , both available now on Hugging Face. ## One Model to Rule ASR, TTS, and Chat The audio model is the more novel of the two. Most Japanese voice pipelines today are assembled from separate components: an ASR model to transcribe speech, a language model to generate a response, and a TTS model to speak it back. LFM2.5-Audio-1.5B-JP is an end-to-end multimodal speech and text language model that does not require separate ASR and TTS components. You speak in, it speaks back , one model, one forward pass. Designed with low latency and real-time conversation in mind, at only 1.5 billion parameters LFM2.5-Audio-JP enables seamless Japanese conversational interaction, achieving capabilities on par with much larger models. That last part is the key claim: the model punches well above its weight class. ## The Architecture Under the Hood The model consists of a pretrained LFM2.5 model as its multimodal backbone, along with a FastConformer-based audio encoder to handle continuous audio inputs, and an RQ-transformer generating discrete tokens coupled with a lightweight audio detokenizer for audio output. FastConformer is a convolution-augmented transformer architecture originally developed by NVIDIA for efficient speech encoding. The RQ-transformer (Residual Quantization transformer) handles converting the language model's outputs into discrete audio codes, which are then decoded into a waveform. The model supports two generation modes, each suited to different tasks: - **Interleaved generation** , generates text and audio tokens in parallel, enabling real-time conversational latency. Best for voice chatbots. - **Sequential generation** , generates one modality at a time, allowing the model to switch between audio and text output on the fly. Best for standalone ASR or TTS tasks. The model details show a 1.2B LM backbone, a 115M FastConformer audio encoder, a 32,768-token context window, and a vocabulary of 65,536 text tokens plus 2,049×8 audio tokens in bfloat16 precision. ## Beating Models 5x Its Size The benchmark story is compelling. On VoiceBench (a Japanese spoken-language evaluation suite), LFM2.5-Audio-1.5B-JP achieves scores competitive with Qwen2.5-Omni-3B (5.5B parameters) and significantly outperforms J-Moshi and LLM-jp-Moshi-v1, both at 7.7B parameters. That is a 5x parameter advantage for the competing models, and they still lose on most metrics. On ASR (measured by character error rate, where lower is better), LFM2.5-Audio-1.5B-JP achieves a CER of 4.42 on the CommonVoice Japanese test set, outperforming Whisper large-v3 (8.5) and all distil-Whisper variants. The one area where it falls short is the ReazonSpeech held-out test set, where specialized ASR models trained on that corpus retain an edge. ![Scatter plot of average JP benchmark performance vs model size in billions of parameters](https://cdn-uploads.huggingface.co/production/uploads/64618cf05dba83471db2be9b/nhW5KrNVrPIe-3zy-RLLt.png) ## The Language Model Gets a Tune-Up Too LFM2.5-1.2B-JP-202606 is Liquid AI's latest general-purpose Japanese chat model, delivering significant improvements in knowledge, instruction following, math, code, and tool use over both comparable-size models and the previous LFM2.5-1.2B-JP. The previous version already beat Qwen3-1.7B and Llama 3.2 1B on JMMLU, M-IFEval, and GSM8K. This update extends that lead across a broader 12-benchmark suite. The numbers are striking. On the full benchmark suite, LFM2.5-1.2B-JP-202606 achieves a domain average of 53.11, compared to 44.83 for Qwen3-1.7B (Instruct) and 39.68 for LFM2.5-1.2B-Instruct. The gap is especially wide on instruction following (66.93 vs 38.45 for Qwen3) and math (62.50 vs 51.20). ![Comprehensive benchmark comparison of LFM2.5-1.2B-JP-202606 against sub-2B models across 12 Japanese benchmarks](https://cdn-uploads.huggingface.co/production/uploads/64618cf05dba83471db2be9b/7gqajAlXAh52nAz85JoQt.png) The model has 1.17B parameters, 16 layers (10 double-gated LIV convolution blocks plus 6 GQA blocks), was trained on 31.5T tokens, and supports a 32,768-token context window. The hybrid convolution-attention architecture is Liquid AI's signature design , it replaces some of the standard attention layers with structured state-space-like convolutions, which improves inference efficiency without sacrificing quality. ## What It's Good For (and What It's Not) The model is recommended for agentic workflows, tool use, structured outputs, bilingual English-Japanese assistants, and on-device personal-assistant applications. It is not recommended for knowledge-intensive tasks. The knowledge cutoff is mid-2024, so anything requiring up-to-date factual recall will hit a wall. Practical use-cases for the pair of models: - **Voice assistants for Japanese markets** , the audio model handles the full speech-in, speech-out loop without stitching together separate services. - **Japanese-language agents** , the text model's strong tool-use and instruction-following scores make it a solid backbone for agentic pipelines. - **On-device deployment** , both models are small enough to run on consumer hardware; the text model is available in GGUF, ONNX, and MLX formats. - **Fine-tuning base** , the audio model is explicitly released as a base for further adaptation, with a documented finetuning interface via the `liquid-audio` package. ## Getting Started The text model plugs straight into the standard Hugging Face Transformers pipeline. The audio model requires Liquid AI's own `liquid-audio` package: ``` pip install liquid-audio pip install flash-attn --no-build-isolation # optional, for FlashAttention 2 ``` For the text model, inference is standard Transformers with a ChatML-style prompt format. It also works with vLLM, SGLang, llama.cpp, and LM Studio out of the box. The model performs best when given clear, explicit instructions that define the task, expected behavior, and output format. Both models are released under the [LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-JP/blob/main/LICENSE). The broader context here is that Liquid AI has been steadily building out a full Japanese-language stack , including a dedicated EN-JP translation nano model and a Japanese PII extraction model , positioning itself as a serious contender for Japanese enterprise AI deployments where data sovereignty and on-device inference matter most. These two releases are the most capable additions to that stack yet. ## Sources - [X post by @liquidai](https://x.com/liquidai/status/2063095523318415744) - [LFM2.5-Audio-1.5B-JP — Hugging Face Model Card](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-JP) - [LFM2.5-1.2B-JP-202606 — Hugging Face Model Card](https://huggingface.co/LiquidAI/LFM2.5-1.2B-JP-202606) - [Liquid AI Complete Model Library — Official Docs](https://docs.liquid.ai/lfm/models/complete-library)