{"slug": "liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b", "title": "Liquid AI Releases LFM2.5-8B-A1B: An On-Device MoE Model With 8.3B Total and 1.5B Active Parameters", "summary": "Liquid AI released LFM2.5-8B-A1B, an on-device Mixture-of-Experts model with 8.3 billion total parameters and 1.5 billion active parameters per token, designed for tool calling and reasoning on consumer hardware. The model features a 128,000-token context window, supports nine languages, and achieved significant benchmark improvements over its predecessor, including a jump in non-hallucination rate from 7.46 to 63.47 and an IFEval score increase from 79.44 to 91.84. The release enables efficient on-device deployment across CPUs, GPUs, and mobile platforms, with inference speeds reaching 253 tokens per second on an M5 Max processor.", "body_md": "Liquid AI just shipped **LFM2.5-8B-A1B**. It is an on-device Mixture-of-Experts (MoE) model built for tool calling. The model holds 8.3B total parameters but activates only 1.5B per token. That sparsity is what lets it run on consumer hardware.\n\nThe release follows LFM2-8B-A1B, which Liquid AI team published earlier. LFM2.5 is a new family of hybrid models for on-device deployment. This version adds a 128K context window, reasoning, and scaled-up training.\n\n**What is LFM2.5-8B-A1B **\n\nThe model uses a sparse MoE design. It activates 1.5B of 8.3B total parameters per forward pass. That keeps each generated token cheap to compute.\n\nThe architecture has 24 layers. Eighteen are double-gated LIV convolution blocks; six are GQA layers. It combines MoE, GQA, and gated short convolution blocks. The context length is 131,072 tokens. The model covers nine languages, including Arabic, Chinese, and Japanese.\n\nLiquid AI team recommends a temperature of 0.2, top_k of 80, and repetition_penalty of 1.05.\n\nUnlike its predecessor, LFM2.5-8B-A1B is a reasoning-only model. It produces an explicit chain of thought before its final answer. Liquid AI team chose this because MoE models run in compute-bound settings. A smaller active parameter count makes each reasoning token inexpensive.\n\n**What Changed Since LFM2-8B-A1B**\n\nLiquid expanded the context window from 32,768 to 128,000 tokens. Pretraining scaled from 12T to 38T tokens. The vocabulary doubled from 65,536 to 128,000 tokens.\n\nThe larger vocabulary tokenizes non-Latin scripts more efficiently. Liquid AI team reports the strongest compression gains in Hindi, Thai, Vietnamese, Indonesian, and Arabic. The rest of the architecture stays the same as LFM2-8B-A1B.\n\n**How Liquid AI Trained It**\n\nLiquid AI team extended the tokenizer in place rather than retraining from scratch. It continued BPE merge training from the original merges on a multilingual corpus. New embedding rows initialize as the mean of their sub-token decompositions. A brief two-stage adaptation then recovers quality.\n\nContext extension came in two phases. A 2T token midtraining phase reached 32K, focused on reasoning, math, and tool use. Raising the RoPE base θ, plus a 400B token stage, reached 128K.\n\nTwo reinforcement learning stages target known failure modes. A preference optimization stage reduces ‘doom loops’ in long reasoning traces. It redistributes probability mass toward plausible alternatives. A separate RL shaping reward discourages loop-inducing restart words like ‘Wait…’. Another RL stage uses an avg@k-based reward to cut hallucinations. The goal is abstention on queries beyond reliable knowledge.\n\n**The Benchmark Case**\n\nLFM2.5-8B-A1B improves over its predecessor across the board. The AA-Omniscience Non-Hallucination Rate jumped from 7.46 to 63.47. IFEval rose from 79.44 to 91.84. MATH500 climbed from 74.80 to 88.76. Tau² Telecom rose from 13.60 to 88.07.\n\nLiquid AI team compared the model against dense and MoE alternatives. On instruction following, it matches Gemma-4-26B-A4B-IT on IFEval. It does so at a fraction of the active parameter count. On Tau² Telecom, it scores 88.07, ahead of much larger models.\n\nThe avg@k reward drives a much lower hallucination rate. Accuracy stays reasonable for the model’s size. On agentic benchmarks, it remains competitive with bigger models.\n\n| Benchmark | LFM2-8B-A1B | LFM2.5-8B-A1B | Δ |\n|---|---|---|---|\n| AA-Omniscience Non-Hallucination Rate | 7.46 | 63.47 | +56.01 |\n| IFEval | 79.44 | 91.84 | +12.40 |\n| MATH500 | 74.80 | 88.76 | +13.96 |\n| Tau² Telecom | 13.60 | 88.07 | +74.47 |\n\n**Running It: CPU, GPU, and Tooling**\n\nThe model ships with day-one support across the inference ecosystem. Frameworks include llama.cpp, MLX, vLLM, and SGLang. ONNX and Liquid’s LEAP edge platform are also supported.\n\nOn CPU, it decodes 253 tokens/s on an M5 Max. It reaches 146 tokens/s on a Ryzen AI Max+ 395. It stays under 6 GB of memory throughout. On a phone, it holds about 30 tokens/s.\n\nOn a single NVIDIA H100 SXM5, output throughput hits 18.5K tokens per second. That is over 1.6B tokens per day at high concurrency.\n\nFor tool use, LFM2.5 writes Pythonic function calls by default. They appear between the `<|tool_call_start|>`\n\nand `<|tool_call_end|>`\n\nspecial tokens. You can override this to JSON in the system prompt.\n\n**Strengths ****and What to Watch**\n\n**and What to Watch**\n\n**Strengths:**\n\n- Activates only 1.5B parameters, keeping inference cheap on edge hardware\n- Competitive instruction-following and agentic scores for its size class\n- 128K context window and nine-language coverage\n- Open-weight under the LFM1.0 license, with base and post-trained checkpoints\n\n**What to Watch**:\n\n**What to Watch**\n\n- Limited knowledge capacity from the small active parameter count\n- Not a fit for heavy programming or knowledge-intensive QA without retrieval\n- Reasoning-only output adds chain-of-thought tokens to every turn\n- Text-only; this variant has no vision or audio input\n\n**Marktechpost’s Visual Explainer**\n\n**01**/ 08\n\n**Key Takeaways**\n\n- Liquid AI's LFM2.5-8B-A1B holds 8.3B total parameters but activates only 1.5B per token.\n- It is reasoning-only, with a 128K context window and nine-language coverage.\n- Non-Hallucination Rate jumped from 7.46 to 63.47 over LFM2-8B-A1B; IFEval reached 91.84.\n- It decodes 253 tok/s on an M5 Max under 6 GB, and ~30 tok/s on a phone.\n- Day-one support spans llama.cpp, MLX, vLLM, and SGLang, with open base and post-trained weights.\n\nCheck out the ** Model Weights **and\n\n**Also, feel free to follow us on**\n\n[Technical details](https://www.liquid.ai/blog/lfm2-5-8b-a1b).**and don’t forget to join our**[Twitter](https://x.com/intent/follow?screen_name=marktechpost)\n\n**and Subscribe to**\n\n[150k+ ML SubReddit](https://www.reddit.com/r/machinelearningnews/)**. Wait! are you on telegram?**\n\n[our Newsletter](https://www.aidevsignals.com/)\n\n[now you can join us on telegram as well.](https://t.me/machinelearningresearchnews)Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? [Connect with us](https://forms.gle/wbash1wF6efRj8G58)", "url": "https://wpnews.pro/news/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b", "canonical_source": "https://www.marktechpost.com/2026/05/28/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b-active-parameters/", "published_at": "2026-05-28 23:29:49+00:00", "updated_at": "2026-05-29 08:12:18.743888+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "machine-learning", "ai-products", "ai-research"], "entities": ["Liquid AI", "LFM2.5-8B-A1B", "LFM2-8B-A1B"], "alternates": {"html": "https://wpnews.pro/news/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b", "markdown": "https://wpnews.pro/news/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b.md", "text": "https://wpnews.pro/news/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b.txt", "jsonld": "https://wpnews.pro/news/liquid-ai-releases-lfm2-5-8b-a1b-an-on-device-moe-model-with-8-3b-total-and-1-5b.jsonld"}}