Liquid AI release LFM2.5-2.6BIt is an agentic system that operates entirely within the device. On phones, laptops or PCs it can plan, call tools, work through multiple-step tasks and even complete them. It has 2,69B total parameter values, a token context window of 131 072 tokens, and 128,000 tokens in its vocabulary. The pre-training required approximately 34 trillion tokens. There are two checkpoints: LFM2.5-2.6B-Base Fine-tuning is required for the best results. LFM2.5-2.6B Post-trained agentic workloads. Data never leaves the device, and marginal costs of every run are near zero. Liquid AI reports tool-use and instruction-following scores competitive with models nearly four times its size.
Is it deployable
Both checkpoints are public on Hugging Face under the same name. Hugging Face has both public checkpoints. lfm1.0 license. The native language weights are a part of the license. GGUF, MLX” ONNX Formats, including vLLM and LM Studio.
- What companies are there?Developers and startup companies can test the model on their own hardware. Model decodes 220 tokens/s in 2.5 GB on M5 Max. Teams in the mid-market can host themselves on a single GPU. NVIDIA H100 SXM5 Around 1.3B tokens are served per day. GGUF, ONNX and GGUF can be used by enterprises and OEMs to push weights across device fleets. LoRA allows for fine-tuning. TRL The following are some examples of how to get started: Unsloth.
- What industries are there?: Liquid AI targets automotive, consumer electronics, industrial robotics, healthcare, financial services, e-commerce” defense. Since no third-party API is contacted, settings which are restricted and “air-gapped” will be the ones that benefit.
- AppsLiquid AI offers recommendations on agentic workloads and tool usage, as well as RAG workflows in long-context. Liquid AI’s practical builds feature on-device assistance, offline document triage with 128K inputs (including invoice and form extraction), robotics command parsing and background agents running continuously, without token cost. Liquid AI does not endorse the model as a solution for knowledge-intensive tasks or agentic coding.
Budget for Architecture and Training
LFM2.5 – 2.6B contains 2.69B parameters total across 30 layers. This stack consists of 22 short double-gated convolution blocks and 8 attention-grouping blocks. Context length and vocabulary size are 131 072 tokens. Tokens used in pre-training were approximately 34 trillion.
Liquid AI increased the vocabulary by doubling it to 128K. extending the existing tokenizer in place It is preferable to start over. The context is extended to 128K by a dedicated phase of mid-training. Model is only text and includes 16 languages.
Four-stage post-training
Four stages are required to make the base checkpoint an agent.
- The first two rounds of supervised fine tuning are performed with a mix that is roughly seven-times the size as used for LFM2.5-8BA1B.
- The second is teacher specialization. One expert in each domain trained using reinforcement learning and verifiable incentives.
- Thirdly, the multi-domain distillation of policy, in which each domain is assigned its own policy, and every prompt route leads to its respective teacher.
- Fourth, agentic learning using GRPO with real harnesses Hermes Agent The following are some examples of how to get started: OpenClaw.
Benchmarks
Liquid AI has compared LFM2.5-2.66B with gemma-4 E2B-it (5.1), gemma-4 E4B-it (8.3), Qwen3.5-4B 4.7B and Qwen3.5-9B 9.7B.
| Benchmark | LFM2.5-2.6B | gemma-4-E4B-it | Qwen3.5-9B |
|---|---|---|---|
| ToolSandbox | 77.83 | 65.00 | 76.44 |
| Multi-IF | 80.07 | 77.35 | 62.55 |
| Its Structure | 85.49 | 76.65 | 78.50 |
| IFBench | 59.17 | 39.24 | 56.47 |
| BFCLv4 | 56.88 | 46.39 | 60.13 |
It leads every instruction-following benchmark reported and nearly every tool use benchmark, trailing Qwen3.5-9B only on BFCLv4. The larger models have an advantage in coding: LiveCodeBenchv6 scores 59.41 against 69.86 on Qwen3.5-9B.
Interactive explainer
The Key Takeaways
- 1.68B params (22 short-conv layers + 8 GQA), 34T context tokens.
- Compares to Qwen3.5-9B and gemma-4E4B it on ToolSandbox Multi-IF, IFStruct and ToolSandbox.
- 220 tok/s with M5 Max. 30 tok/s in phones under 2.5 GB.
- Open weights under
lfm1.0GGUF MLX ONNX is available from the first day.
Take a look at the Technical details, LFM2.5-2.6B” LFM2.5-2.6B-Base. Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe Now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.
Want to promote your GitHub repo, Hugging Face page, Product release or Webinar?? Connect with us
Asif Razzaq, CEO of Marktechpost Media Inc. is a visionary engineer and entrepreneur who is dedicated to harnessing Artificial Intelligence’s potential for the social good. Marktechpost was his most recent venture. This platform, specializing in Artificial Intelligence, is known for its comprehensive coverage of deep learning and machine learning. Over 2 million views per month are a testament to the platform’s popularity.

