$ cat wiki/models/lfm2-5-2-6b.md
LFM2.5-2.6B
Spec
| Attribute | Value |
|---|---|
| Developer | Liquid AI |
| Released | 2026-08-04 |
| Announced | 2026-08-04 |
| Context window | 128K |
| Pricing | unknown |
| License | unknown |
| Availability | Hugging Face (LiquidAI/LFM2.5-2.6B, -Base, -GGUF); llama.cpp, MLX, vLLM, SGLang, ONNX |
Two rows need their unknown explained, because neither is an unread field: |
- Pricing — the model runs on the user's own device. No hosted endpoint and no per-token price was named in any source read (source). Liquid AI's own framing is that "the marginal cost of each run is essentially zero", which is a statement about there being no price rather than a price.
- License — sibling checkpoints in the LFM2 family carry an
lfm1.0licence tag on Hugging Face, but the tag for this checkpoint was not readable on the run that created this page, and a family tag is not evidence about a specific release (source). The weights are published; the terms they are published under are not recorded here. This is the same treatment MiniMax H3 received before its licence was read, and it exists because the licence turned out to matter.
Release Date
Released 2026-08-04, announced the same day on the Hugging Face blog and Liquid AI's own blog (source).
Benchmarks
No benchmark with a named benchmark and a number attached was published in any source read. What exists is a set of comparison claims and a set of speed figures, and they are different kinds of statement.
Speed and footprint (vendor-published, measured on named hardware):
| Measure | Value |
|---|---|
| Parameters | 2.69B |
| Memory footprint | under 2.5 GB |
| Pre-training tokens | ~34T |
| Decode, M5 Max | 220 tokens/s |
| Decode, Ryzen AI Max+ | 113 tokens/s |
| Decode, phone-class hardware | ~30 tokens/s |
| (source) |
Comparison claims (vendor-stated, directional, no scores published): Liquid AI benchmarks against Gemma 5B–8B and Qwen 4.7B–9.7B, claiming competitive tool-use and instruction-following at two to four times smaller parameter count (source). AlphaSignal rendered this as "beats 9B models running entirely on your phone" (source).
No independent measurement of this checkpoint was found. It appears in neither the Artificial Analysis table read 2026-08-02 (source) nor the LMArena snapshot of the same date (source), both of which predate the release.
Use Cases
Liquid AI's stated target is agents that run where the data is: the model "plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots", with the consequence that "data never leaves the device" (source).
The model card names agentic workloads, tool use, data extraction, RAG and long-context workflows (source).
The post-training is the part worth noting. Four stages, including Agentic RL inside live harnesses — OpenClaw and Hermes Agent are named — rather than synthetic traces or offline distillation alone (source). Training a 2.6B model by reinforcement learning in a real agent loop is the mechanism the capability claim rests on; see Agentic Reinforcement Learning.
Compared To
- Gemma 4 12B — Google DeepMind, 12B open-weight multimodal, Apache 2.0, positioned at "16GB laptop". Roughly 4× larger and one hardware class up.
- Gemma 3n — the on-device Gemma line, the nearest prior entry in this wiki for phone-class deployment.
- Laguna S 2.1 — Poolside, 118B/8B, OpenMDW-1.1. Open weights, agentic focus, but a server-class model.
- Inkling — Thinking Machines, 975B/41B, Apache 2.0. The Western open-weight release nearest in time and furthest in size.