What LFM2.5-2.6B is
LFM2.5-2.6B is an on-device language model from Liquid AI. It builds on the company's own LFM2 architecture, adding a 131,072-token context window and agentic post-training.
The specifications:
| Item | Value |
|---|---|
| Total parameters | 2.69B |
| Layers | 30 (22 double-gated short convolution blocks + 8 GQA — grouped attention that shares computation to stay light) |
| Training budget | 34 trillion tokens |
| Vocabulary size | 128,000 |
| Context length | 131,072 tokens |
| Languages | 16 |
Several distribution formats are available: the base model plus GGUF for CPU execution, ONNX for portability, and MLX for Apple Silicon, so you can match it to the hardware you have.
One more trait worth knowing: it is a reasoning model that always thinks before answering, so a thinking tag is inserted automatically at the start of a response. Agentic work means assembling JSON to pass to tools over and over, so having a way to format and inspect JSON locally speeds up the verification loop.
It builds on the LFM2 architecture with a 128K context window and agentic post-training. ／ Total parameters<strong>: 2.69B ／ Number of layers</strong>: 30 (22 double-gated short convolution blocks + 8 GQA) ／ Training budget<strong>: 34 trillion tokens ／ Vocabulary size</strong>: 128,000 ／ Context length**: 131,072 tokens ／ English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish ／ LFM2.5-2.6B is a pure reasoning model that always thinks before it answers. — From the passages on the architecture, context length and post-training, the specification items, the list of supported languages, and the reasoning behaviour
What "competitive with models 4x larger" actually covers
The claim Liquid AI pushes hardest concerns performance when the model is given tools. On tool calling, instruction following and multi-step agentic tasks, it says the model is competitive with models four times its size.
The reasoning it offers is the training method: reinforcement learning conducted inside the most popular agentic harnesses, to improve compatibility with the setups people actually use. The argument is that it was trained where it runs, not for a benchmark.
The published comparison table sets it against sub-10B models. On BFCLv4, a tool-use evaluation, LFM2.5-2.6B scores 56.88 against 46.39 for gemma-4-E4B-it (8B), roughly three times its parameter count. On IFBench, for instruction following, the figures are 59.17 against 39.24.
The numbers do not win everywhere. In the same table LiveCodeBenchv6 comes in at 59.41, below gemma-4-E4B-it (8B) at 63.77. The self-reported weakness at coding and the table agree with each other.
Competitive with models 4x larger on tool use, instruction following, and multi-step agentic tasks. ／ Trained inside the most popular agentic harnesses to improve compatibility. ／ We compared LFM2.5-2.6B with relevant sub-10B models on a diverse suite of benchmarks. ／ | BFCLv4 | 56.88 | 36.98 | 46.39 | 50.56 | 60.13 | ／ | IFBench | 59.17 | 34.08 | 39.24 | 48.40 | 56.47 | ／ | LiveCodeBenchv6 | 59.41 | 54.92 | 63.77 | 60.85 | 69.86 | — From the claim of competitiveness with 4x-larger models, the training inside agentic harnesses, the comparison set, and the three benchmark rows cited above
Speed and memory: the case for running it on a phone
For anything calling itself on-device, speed and memory are what count. Liquid AI reports it as the fastest model in its own testing, with decode speeds of 220 tokens per second on an Apple M5 Max and 113 on a Ryzen AI Max+ 395, in under 2.5GB of memory.
The phone claim is specific. The model card says that at 30 tokens per second you can run capable agents even on a phone. The figures above were measured on high-end PC chips and will not appear as-is on a handset. But with the required line stated as 30 tokens/sec, you can estimate whether your device clears it.
There are server-side numbers too: close to 15,000 output tokens per second at high concurrency on GPUs, roughly 1.3 billion tokens a day on a single H100. Being able to use the same model locally and on a server is an operational advantage in itself.
220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU, in under 2.5 GB of memory. ／ Due to its efficient LFM2 architecture, LFM2.5-2.6B is the fastest model we tested, with decode speeds of 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395. ／ At 30 tokens/s, it allows you to run capable agents even on a phone. ／ … reaching almost <strong>15K output tokens per second at high concurrency</strong>, roughly 1.3B tokens per day on a single H100 — From the passages on decode speed and memory use, the chips used for measurement, the practical line stated for phones, and GPU throughput
Choosing between this and Maple-Preview
Both are small models that run locally, but Maple-Preview is aimed in the opposite direction.
| Aspect | LFM2.5-2.6B | Maple-Preview |
|---|---|---|
| Strengths | Tool use, agentic work, data extraction, RAG | Pure reasoning; Olympiad-level maths problems |
| Stated weaknesses | Agentic coding, knowledge-heavy tasks | Agentic benchmarks (minimal post-training) |
| Size | 2.69B | 20B-A1B (8 experts active) |
| Licence | Bespoke lfm1.0 | MIT |
| Claimed speed | 220 tokens/sec on an M5 Max | 218 tokens/sec on an M4 Mac mini |
LFM2.5 if you want to run agents; Maple-Preview if you want hard problems solved. The headline speeds are almost identical, but the work you hand over changes the outcome entirely.
For commercial use, the licence difference matters. Maple-Preview is MIT; LFM2.5-2.6B carries a bespoke lfm1.0, classified on Hugging Face as "other". It is not a widely known permissive licence, so the terms have to be read in the licence text itself.
license: other ／ license_name: lfm1.0 ／ We recommend using it for agentic workloads, tool use, data extraction, RAG, and long-context workflows. It is not recommended for agentic coding and knowledge-heavy tasks. — From the licence classification and name, and the recommended and non-recommended use cases
Today we introduce Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM. ／ It solves IMO-level problems ／ It utilizes a 24-layer, 256-expert (8 active) configuration ／ 218 tok/s M4 Mac mini ／ may underperform on agentic benchmarks ／ This preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning. ／ Maple-Preview is released under the [MIT License](LICENSE). — From the passages behind the Maple-Preview column: size, strengths, active experts, speed, weakness and licence
How to read LFM2.5-2.6B
The centre of LFM2.5-2.6B is agents that run locally. At 2.69B it fits in under 2.5GB of memory, and on tool-use evaluations it posts numbers above models larger than itself.
The areas Liquid AI marks as not recommended are best taken at face value. This is not where you send coding work, and it is not suited to questions that require knowing a lot. Confined to calling tools and moving through steps, this size holds up perfectly well.
If you are evaluating it, start by separating the work into "using tools" and "answering from knowledge". The former points to this model; the latter points elsewhere. To keep the local options in view, following open-weight model releases alongside it gives you more to compare against.
