← Back to feed

LiquidAI published LFM2.5-2.6B to Hugging Face with benchmarks showing it beats every compared model up to approximately 10 billion parameters on tool use, instruction following, and multi-step agentic tasks. The model is available on Ollama and Hugging Face with no access restrictions.

What the Numbers Show

LiquidAI evaluated LFM2.5-2.6B against a group of models up to four times its size on STEM reasoning, instruction following, tool use, and agentic task completion. On tool use and instruction following, the 2.6B model leads across the board. The gains are not marginal: LiquidAI describes the 2.6B as “best-in-class” on agentic capability within the sub-10B range.

The prior LFM2.5 generation was an 8B model, trained on 38 trillion tokens. The 2.6B variant is a materially smaller architecture, not a quantised version of the 8B. LiquidAI has not published training token counts or architecture details for the 2.6B.

Architecture and Tool Interface

LFM2.5-2.6B has two properties that distinguish it from comparably-sized transformer models:

Native tool calling: Function calls use a Pythonic syntax wrapped in <|tool_call_start|> and <|tool_call_end|> tokens. This is a structured output mechanism built into the model rather than implemented via prompt injection — a design that reduces hallucinated tool calls compared to instruction-tuned models that approximate tool use through text.

Built-in reasoning: The model has a reasoning phase between <think> and </think> tags before producing its final response. At 2.6B parameters, on-device reasoning before committing to an answer is a genuine capability differentiator. Models in this size range typically do not have reliable chain-of-thought behaviour.

Deployment Implications

At 2.6B parameters, LFM2.5-2.6B runs on consumer hardware. An iPhone 16 Pro can run a 2.6B model comfortably at usable token rates. The combination of native tool calling and reasoning at this scale opens a deployment pattern that has been constrained: a locally-running agent that can call APIs, follow multi-step instructions, and reason about its own outputs — without cloud inference costs or latency.

The model is available through ollama pull lfm2.5, which means it is immediately accessible to any developer already running Ollama for local inference. There is no waiting list or API key requirement.

The Larger LFM2.5 Family

The 2.6B is the smallest model in the LFM2.5 line. LiquidAI has previously published LFM2.5 at 8B, and the 8B model’s benchmark profile — hallucinations cut 9x versus the prior generation — established LiquidAI as a serious competitor in the sub-frontier open-weight category. The 2.6B extends that work into a size range previously dominated by quantised versions of Qwen or Llama models, neither of which was designed natively for agentic use.

For enterprises evaluating on-device or edge agent deployments — point-of-sale, industrial control, medical device — the LFM2.5-2.6B adds a credible option with a commercial-friendly release posture.