LFM2.5-2.6B

LFM2.5-2.6B

Quick answer: LFM2.5-2.6B is Liquid AI's 2.6B-parameter (2.69B total) on-device agentic model, released August 4, 2026, built on the LFM2 hybrid architecture with a 131,072-token context window. Trained with agentic reinforcement learning inside real harnesses (Hermes Agent, OpenClaw), it scores 59.2% on IFBench, 80.1% on Multi-IF, and 77.8% on ToolSandbox — competitive with, and often ahead of, models roughly 4x its size. It decodes at 220 tokens/s on an Apple M5 Max and 113 tokens/s on a Ryzen AI Max+ 395, under 2.5 GB of memory. Open weights under Liquid's lfm1.0 license.

At a Glance

Where LFM2.5-2.6B leads

• 80.1% Multi-IF and 85.5% IFStruct — leads every instruction-following benchmark tested, ahead of models up to 4x its size
• 77.8% ToolSandbox and 62.9% Claw-Eval — leading or near-leading tool-use and agentic scores at 2.6B parameters
• Fastest model in its class: 220 tok/s on Apple M5 Max, ~15K output tok/s at high GPU concurrency (single H100)
• Under 2.5 GB memory footprint — runs on a phone at 30 tok/s

Where it lags

• 56.9% BFCLv4 — trails Qwen3.5-9B (60.1%), the largest model in its comparison set
• 59.4% LiveCodeBenchv6 — coding is the one area where larger comparison models (Qwen3.5-9B at 69.9%) keep a clear edge
• Not recommended by Liquid AI for agentic coding or knowledge-heavy tasks

Best for: on-device / edge agentic workloads — tool use, data extraction, RAG, and long-context agent tasks where local, private, zero-marginal-cost inference matters more than raw coding or knowledge-recall performance.

What LFM2.5-2.6B Is

LFM2.5-2.6B is Liquid AI's latest small on-device model, purpose-built to run agentic workflows — planning, tool calls, multi-step tasks — entirely locally, without a cloud API. It's built on the LFM2 hybrid architecture (22 double-gated short-convolution blocks + 8 GQA attention blocks across 30 layers), pre-trained on ~34 trillion tokens, with a dedicated mid-training phase extending context to 128K tokens and a doubled 128K-token vocabulary to better support non-Latin scripts.

Post-training turns the base model into an agent through a four-stage pipeline: two rounds of supervised fine-tuning, per-domain teacher specialization with reinforcement learning from verifiable rewards, multi-domain on-policy distillation (where a single student model is supervised by domain-expert teachers under its own rollout policy), and a final agentic reinforcement-learning stage run directly inside real agent harnesses (Hermes Agent, OpenClaw, and others). That last stage is Liquid's key differentiator: by training inside actual harnesses rather than on synthetic tool-call data, the model is exposed to real tool schemas, system prompts, and interaction patterns, which Liquid credits for its outsized tool-use and instruction-following scores relative to its 2.6B size.

On Benchgen's read, LFM2.5-2.6B's benchmark profile is unusually lopsided in a good way for an edge model: it leads comparably-sized-and-larger models on every instruction-following and most tool-use/agentic benchmarks tested, while trailing on raw coding (LiveCodeBenchv6) and the largest comparison model on BFCLv4 function-calling — consistent with Liquid's own guidance that it's best suited to agentic, tool-use, and long-context workloads rather than coding-heavy or deep-knowledge tasks.

Specifications

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face in native (Transformers/vLLM/SGLang), GGUF (llama.cpp), ONNX, and MLX (Apple Silicon) formats. Designed for zero-marginal-cost local inference: decodes at 220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen AI Max+ 395 CPU, or ~15,000 output tok/s at high concurrency on a single NVIDIA H100 GPU (~1.3B tokens/day).

Public Benchmark Scores

Scores are self-reported by Liquid AI in the official LFM2.5-2.6B launch post, evaluated with vLLM under fixed generation settings per benchmark (documented in the blog's methodology footnote).

Notable In-House / Unlisted Benchmark Results

Liquid AI also reports results on a few evaluations not yet tracked as standalone Benchgen leaderboard entries, either due to unclear public provenance or a scale that couldn't be reliably cross-verified against existing Benchgen benchmark pages:

• IFStruct — 85.49%, Liquid AI's own structured-instruction-following eval; not independently verifiable as a standalone public benchmark.
• PinchBench — 68.22%; no independently verifiable public benchmark definition found.
• τ³-Bench Banking — 5.67 (Liquid AI's reported scale); Benchgen's existing τ³ Banking page uses a "% tasks completed" metric on a 20–33% scale for other models, which doesn't reconcile with Liquid's reported numbers, so this score is not listed as a leaderboard entry pending clarification of methodology.
• BrowseComp+ (OpenClaw) — 26.89; an OpenClaw-harness-specific variant of BrowseComp, not confirmed equivalent to Benchgen's existing BrowseComp benchmark.

LFM2.5-2.6B vs Alternatives

Despite being the smallest model in the comparison — roughly 2–4x smaller than the alternatives — LFM2.5-2.6B leads on every instruction-following benchmark (IFBench, Multi-IF) and nearly every tool-use benchmark, trailing only Qwen3.5-9B on BFCLv4 function calling. All comparison scores are as reported by Liquid AI in the same launch benchmark table; see each model's own page for scores from other sources.

FAQ

Is LFM2.5-2.6B open source? Yes — both the base (LFM2.5-2.6B-Base) and post-trained (LFM2.5-2.6B) checkpoints are available on Hugging Face under Liquid AI's lfm1.0 license, downloadable, fine-tunable, and deployable.

How big is LFM2.5-2.6B? 2.69 billion total parameters (dense, not mixture-of-experts), built on the LFM2 hybrid architecture (short-convolution + GQA attention blocks).

What is LFM2.5-2.6B's context window? 131,072 tokens (128K), reached via a dedicated context-extension phase during mid-training.

How fast is LFM2.5-2.6B? 220 tokens/s on an Apple M5 Max and 113 tokens/s on an AMD Ryzen AI Max+ 395 CPU, under 2.5 GB of memory — fast enough to run a capable agent on a phone at ~30 tokens/s. On GPU, it reaches almost 15,000 output tokens/s at high concurrency on a single NVIDIA H100.

What is LFM2.5-2.6B best used for? Liquid AI recommends it for agentic workloads, tool use, data extraction, RAG, and long-context workflows on edge devices — and explicitly does not recommend it for agentic coding or knowledge-heavy tasks, where larger models still lead.

Benchmark scores sourced from Liquid AI's official LFM2.5-2.6B launch announcement (liquid.ai/blog, Aug 4, 2026) and the LiquidAI/LFM2.5-2.6B model card on Hugging Face. Comparison-model scores (Gemma 4 E2B/E4B, Qwen3.5-4B/9B) are as reported in that same table.