LFM2.5-2.6B
Liquid AI's 2.6B open-weight model tuned for on-device agents — tool calling and multi-step planning that runs locally on phones and laptops.
Updated 2026-08-10
Yes. LFM2.5-2.6B is free to use.
Listed pricing: Free open weights — self-host only, no hosted API.
Overview
LFM2.5-2.6B is Liquid AI's August 2026 open-weight release: a 2.6-billion-parameter model trained for agentic work — planning a task, calling a tool, reading the result, taking the next step — at a size that fits on a phone, a laptop or an edge box rather than a GPU server. It sits at the top of the LFM2 family, which spans sub-billion checkpoints up through this one, and it ships as weights you download and run yourself. There is no hosted endpoint and no per-token bill.
The audience is developers building local agents: offline assistants, on-device automation, and privacy-constrained pipelines in health, legal or finance where shipping user data to a hosted API is the thing that kills the project. What separates it from the average open small model is the training target. Most 2–3B checkpoints are optimised for chat and completion with function calling bolted on afterwards; Liquid positions tool calling and multi-step execution as the primary job. The architecture follows the same logic — the LFM2 line interleaves short convolution blocks with grouped-query attention instead of using a uniform transformer stack, a choice aimed at throughput on CPUs and mobile NPUs rather than leaderboard scores on datacentre GPUs. Liquid also ships a first-party edge path (the LEAP SDK for iOS and Android, plus the Apollo app for running models locally), which is more deployment scaffolding than most labs put behind a Hugging Face repo.
The honest read for 2026: the small-open-model tier is no longer a gap in the market, it is a scrum. Qwen, Gemma, Phi and the SmolLM line all publish tool-calling-capable checkpoints in this weight class, several under Apache 2.0 or MIT. LFM2.5-2.6B's case rests on the mobile deployment story and the architecture's CPU/NPU throughput, not on being the only option. Two caveats before you build on it. First, the performance claims are vendor-published and, this soon after release, largely un-replicated — benchmark it on your own task, not on the launch post. Second, read the licence text yourself: Liquid has released LFM weights under its own LFM Open License, which turns on company revenue rather than granting the unrestricted Apache 2.0 rights its closest rivals do. If you plan to embed this in something you sell, that clause is the first thing to check, not the last.
Is LFM2.5-2.6B free?
Yes. LFM2.5-2.6B is free to use.
What the free tier covers: The entire model is free to download and run — the only cost is the hardware you run it on and the licence terms you agree to.
Listed pricing: Free open weights — self-host only, no hosted API.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
LFM2.5-2.6B pricing
| Plan | Price | What's included |
|---|---|---|
| Open weights | Free | Full model weights for self-hosting on your own hardware. No hosted API, no usage metering, no rate limits. Commercial use is governed by Liquid's own model licence — read the terms on the model card, as the LFM Open License has carried a company-revenue threshold rather than being plain Apache 2.0. |
Full model weights for self-hosting on your own hardware. No hosted API, no usage metering, no rate limits. Commercial use is governed by Liquid's own model licence — read the terms on the model card, as the LFM Open License has carried a company-revenue threshold rather than being plain Apache 2.0.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Is LFM2.5-2.6B worth it?
Worth it for Mobile and edge developers building offline agents for iOS or Android who need dependable tool calling from a model small enough to ship inside the app, with no hosted API in the loop.
You can test that on the free tier before paying anything. The recorded trade-offs are listed below, and any one of them can settle the question on its own.
The 8/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with LFM2.5-2.6B. How we verify.
Worth it if
The strengths recorded against this entry.
- Runs entirely offline on phone- and laptop-class hardware, so there is no per-token cost and no user data leaves the device
- Tuned specifically for tool calling and multi-step execution rather than conversation — the weak spot for most models this size
- Open weights you can fine-tune, quantize and ship inside your own application, with no endpoint a vendor can deprecate
- First-party mobile deployment path (LEAP SDK for iOS/Android, Apollo app) instead of leaving you to port it yourself
- Architecture picked for CPU and NPU throughput, which is the constraint that actually bites on edge hardware
Not worth it if
Any one of these blocks your use case.
- Licence is Liquid's own LFM Open License with a company-revenue threshold, not Apache 2.0 or MIT — a real disadvantage against Qwen, Gemma and SmolLM checkpoints in the same weight class
- No hosted API at all — you own inference, quantization, updates and the ops burden that comes with them
- At 2.6B parameters it will not hold a long-horizon coding session or deep multi-file reasoning; this is a weight class for bounded, tool-mediated tasks
- Vendor-published benchmarks remain largely un-replicated by independent evaluators, so treat the launch numbers as a starting hypothesis rather than a result
What sets LFM2.5-2.6B apart
- Ships a first-party mobile deployment path — the LEAP SDK for iOS/Android and the Apollo local-model app — where most open small models stop at a Hugging Face repo and leave the porting to you
- Hybrid short-convolution plus grouped-query-attention backbone rather than a uniform transformer stack, optimised for CPU and mobile NPU throughput instead of GPU benchmark runs
- Agentic tool calling is the primary training objective, not a capability grafted onto a chat model after the fact
- Licensed under Liquid's own LFM Open License with a company-revenue threshold — a genuine difference from Apache 2.0 rivals in this weight class, and one that counts against it for commercial embedding
Key features
Agentic tool use
Trained for function calling and multi-step task execution rather than chat alone, so the model can plan a sequence, invoke a tool, read the output and continue — the loop most sub-3B open models handle unreliably.
Runs fully on-device
Sized for phones, laptops and edge hardware, which means no inference bill, no network round-trip and no user data leaving the device. An agent built on it keeps working with no connection at all.
First-party mobile deployment stack
Liquid ships the LEAP SDK for iOS and Android plus the Apollo app for running models locally, so getting the weights onto a handset is a supported path rather than a porting project you own.
Hybrid conv-attention backbone
The LFM2 architecture interleaves short convolution blocks with grouped-query attention instead of a uniform transformer stack — a design aimed at CPU and mobile NPU throughput rather than datacentre GPU benchmark scores.
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| LFM2.5-2.6B | Mobile and edge developers building offline agents for iOS or Android who need dependable tool calling from a model small enough to ship inside the app, with no hosted API in the loop. | Free open weights — self-host only, no hosted API | 8/10 |
| Claude Code vs Claude Code → | Developers who want an agent that works inside an existing repo and toolchain rather than in a hosted editor, and who already pay for a Claude plan. | Included with Claude Free, Pro, Max, Team and Enterprise plans | 9.3/10 |
| Cursor vs Cursor → | Professional developers who want an IDE-native agent that can also run in the cloud and be steered from a phone, and engineering teams that need admin control over which model handles which request. | Free Hobby + Individual $20/mo + Teams $40/user/mo + Enterprise custom | 9.2/10 |
| v0 by Vercel vs v0 by Vercel → | Frontend developers on React and Next.js who want production-ready UI generated from a text prompt or a screenshot instead of hand-built boilerplate. | Free tier + Plus $30/user/mo + Business $100/user/mo + Enterprise custom | 9/10 |
Compare head-to-head
Related reading
Claude Riemann Hypothesis Result: 41.6% to 67.2%
Anthropic says an unreleased Claude raised the proven fraction of Riemann zeta zeros on the critical line from 41.6% to 67.2%. Here is what that means.
GPT-5.6-Cyber vs GPT-5.6 Sol: The Daybreak Split
OpenAI's GPT-5.6-Cyber is gated on authorized cybersecurity work. The Daybreak Blue and Red tiers, the 95% claim, and who each model is meant for.
Muse Glimmer and the 30B Open-Weights Ceiling
Meta released Muse Glimmer's 30B weights under Apache 2.0, landing in the same consumer-GPU size class as every other recent open model release.
Ready to try LFM2.5-2.6B?
Head to the official site to start with LFM2.5-2.6B — pricing and plans are listed above.
Visit LFM2.5-2.6B