Liquid AI's LFM2.5-2.6B matches models four times its size
Liquid AI published the Hugging Face model card for LFM2.5-2.6B, the newest entry in its LFM2.5 family of hybrid models built for on-device deployment. It extends the LFM2 architecture with a 128K-token context window and agentic post-training, and Liquid AI calls it "best-in-class" among agents, saying it is competitive with models four times larger on tool use, instruction following, and multi-step agentic tasks.
The spec sheet lists 2.69 billion total parameters (rounded to 2.6B in the name), 30 layers made up of 22 double-gated short convolution blocks plus 8 grouped-query-attention blocks, a 128,000-token vocabulary, and a context length of 131,072 tokens. It was pretrained on 34 trillion tokens and supports 16 languages, including English, Arabic, Chinese, and Russian. Post-training runs in four stages: two rounds of supervised fine-tuning, per-domain teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning, in which Liquid AI trains the model directly inside popular agentic harnesses so it learns their tools, system prompts, and interaction patterns.
On the performance side, Liquid AI reports CPU decode speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, both using under 2.5GB of memory, and calls it the fastest model the company tested on CPU. It adds that 30 tokens per second is enough to run capable agents even on a phone. On GPU, Liquid AI says the model reaches almost 15,000 output tokens per second at high concurrency, translating to roughly 1.3 billion tokens per day on a single H100, and describes it as the fastest model in its size class.
Liquid AI recommends LFM2.5-2.6B for agentic workloads, tool use, data extraction, retrieval-augmented generation, and long-context workflows, but explicitly not for agentic coding or knowledge-heavy tasks. It is a reasoning model that always thinks before answering, wrapping that reasoning in a
The model card does not show the actual benchmark numbers behind the "4x larger" comparison, nor does it name which larger models were tested against. No release date is given beyond the citation year 2026, no licensing terms or pricing are stated, and no individual researchers are credited, only the organization "Liquid AI." The GPU throughput figures also do not specify concurrency level, batch size, or precision.
Key facts
- LFM2.5-2.6B has 2.69 billion total parameters, 30 layers (22 double-gated short convolution blocks plus 8 GQA blocks), a 131,072-token context window, and was pretrained on 34 trillion tokens.
- Liquid AI says the model is competitive with models four times larger on tool use, instruction following, and multi-step agentic tasks, trained via agentic reinforcement learning directly inside popular agentic harnesses.
- On CPU it decodes at 220 tokens/s on an Apple M5 Max and 113 tokens/s on a Ryzen AI Max+ 395 using under 2.5GB of memory; on GPU it reaches almost 15,000 output tokens/s at high concurrency, about 1.3 billion tokens/day on a single H100.
- Liquid AI recommends it for agentic workloads, tool use, data extraction, RAG, and long-context tasks, but not for agentic coding or knowledge-heavy tasks.
- The model had 89,680 downloads on Hugging Face in the past month and scores 85.49 on the Ifstruct V1 evaluation leaderboard.
Why it matters
Most agentic model releases chase raw capability at large parameter counts. LFM2.5-2.6B goes the other direction: Liquid AI is betting that a 2.69B-parameter model, if it is trained specifically for tool use and agent harnesses, can hold its own against models four times its size while running on ordinary consumer CPUs. If that holds up, it lowers the hardware bar for running agents locally instead of routing every step through a cloud API.
Who it affects
Developers building on-device or edge agents, including on laptops and phones, who want to avoid cloud inference costs and latency. It also matters to teams doing tool use, data extraction, RAG, or long-context work who need a small model that still handles multi-step agentic tasks reliably.
How to use it
The model is on Hugging Face and works with Transformers, vLLM, SGLang, and Docker Model Runner, with example code for each in the model card. It supports function calling through a Pythonic syntax by default (JSON calls can be requested via the system prompt) and uses a ChatML-like chat template. Liquid AI recommends it for agentic workloads, tool use, data extraction, RAG, and long-context workflows, and explicitly advises against using it for agentic coding or knowledge-heavy tasks. No licensing terms or pricing are stated in the model card.
How solid is it
The comparison against "4x larger" models is Liquid AI's own claim on its own model card; the underlying benchmark scores and the names of the larger models being compared against are not shown in the published text, and the Ifstruct V1 leaderboard the model scores 85.49 on is Liquid AI's own benchmark. The CPU speed figures (220 and 113 tokens/s) and the GPU throughput figures (near 15,000 tokens/s, ~1.3B tokens/day on an H100) are also self-reported, with the GPU numbers missing concurrency, batch size, and precision details.
Risks and caveats
Everything here is self-published by the model's own developer with no third-party benchmark shown. The '4x larger' claim names no specific competing models. No release date beyond the citation year 2026, no licensing or pricing terms, and no individual researchers are credited, only 'Liquid AI' as an organization. The name '2.6B' also rounds down from the stated 2.69B total parameters, which the model card does not explain.