LFM2.5
modelYour notes
Liquid AI's 2026 on-device generation, built on the LFM2 hybrid backbone (double-gated short convolution blocks interleaved with grouped-query attention) and rolled out as a series of releases from January to August 2026. LFM2.5 is a continued-pretraining + RL refresh of the LFM2 checkpoints rather than a new pretrain: the 1.2B and 350M extend LFM2's 10T-token budget to 28T, the 8B-A1B MoE goes from 12T to 38T, and the 2.6B to 34T; the vocabulary was doubled to 128K by extending the existing BPE merges in place ("rather than retraining the model from scratch") followed by embedding-only and then full-model continued pretraining. Every text release then gets large-scale multi-stage RL. Weights ship under the LFM Open License v1.0 (Liquid's Apache-derived open-weight license), with GGUF / MLX / ONNX exports and day-one llama.cpp, vLLM, SGLang and LEAP support.
Members: LFM2.5-1.2B Base / Instruct / Thinking / JP (Jan 2026, 32K context, 65K vocab), LFM2.5-VL-1.6B and LFM2.5-Audio-1.5B (Jan; native speech I/O with an 8× faster LFM-based audio detokenizer), LFM2.5-350M (Mar), LFM2.5-VL-450M (Apr), LFM2.5-8B-A1B (May; 8.3B total / 1.5B active, 32 experts × top-4, 18 conv + 6 GQA layers, 128K context, reasoning-only with explicit CoT, plus an avg@k abstention reward that cut the AA-Omniscience hallucination penalty from −78 to −25), LFM2.5-230M (Jun; 19T tokens, smallest LFM), Embedding/ColBERT-350M retrievers and Encoder-230M/350M bidirectional encoders (Jun–Jul), LFM2.5-2.6B (Aug; 22 conv + 8 GQA layers, 128K context, four-stage agentic post-training — SFT, teacher specialization, multi-domain on-policy distillation, agentic GRPO inside Hermes Agent / OpenClaw harnesses via verl + SGLang), LFM2.5-VL-3B (Aug; 2.6B backbone + SigLIP2 NaFlex 400M), and DSpark speculative-decoding drafters (Aug; 296–328M, up to 3.2× decode speedup, upstreamed to llama.cpp and SGLang).
LFM2.5-8B-A1B: IFEval 91.84, IFBench 56.47, Multi-IF 79.93, AIME25 42.53, BFCL v3 64.36 / v4 48.50, τ²-Telecom 88.07 — +12 to +74 points over LFM2-8B-A1B. LFM2.5-2.6B: AIME25 51.87, LiveCodeBench v6 59.41, IFBench 59.17, BFCL v4 56.88, ToolSandbox 77.83, beating Gemma-4-E4B and matching Qwen3.5-4B at 220 tok/s on an M5 Max. AA Intelligence Index v4.3: 7 (8B-A1B), 8 (2.6B), 5 (1.2B-Instruct / Thinking) — the two larger models above LFM2-24B-A2B's 6.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| IFEval | 91.84 | 8B-A1B |
| IFBench | 56.47 | 8B-A1B |
| Multi-IF | 79.93 | 8B-A1B |
| MATH-500 | 88.76 | 8B-A1B |
| AIME25 | 42.53 | 8B-A1B |
| BFCL v3 | 64.36 | 8B-A1B |
| BFCL v4 | 48.50 | 8B-A1B |
| τ²-Bench Telecom | 88.07 | 8B-A1B |
| AA-Omniscience Index | -24.70 | 8B-A1B |
Variants
| Name | Parameters | Notes |
|---|---|---|
| LFM2.5-1.2B-Instruct | 1.17B | Jan 5 2026 launch model; 16 layers (10 conv + 6 GQA), 28T tokens (extended from LFM2-1.2B's 10T), 32K context, 65K vocab. GPQA 38.89, IFEval 86.23, IFBench 47.33, BFCLv3 49.12. AA Intelligence Index v4.3 score 5. Base checkpoint LFM2.5-1.2B-Base also released. |
| LFM2.5-1.2B-Thinking | 1.17B | Jan 20 2026; reasoning variant with explicit thinking traces, <900 MB on a phone. GPQA Diamond 37.86, MATH-500 87.96, AIME25 31.73, IFEval 88.42. AA Intelligence Index v4.3 score 5. |
| LFM2.5-1.2B-JP | 1.17B | Japanese-optimized chat model (Jan 2026; refreshed as LFM2.5-1.2B-JP-202606 in June). JMMLU 50.7, M-IFEval-ja 58.1. |
| LFM2.5-VL-1.6B | 1.6B | Jan 2026 vision-language refresh on the LFM2.5-1.2B backbone; multi-image and 7-language vision understanding. MMStar 50.67, MM-IFEval 52.29. VL-1.6B-Extract / VL-450M-Extract OCR variants followed in June. |
| LFM2.5-Audio-1.5B | 1.5B | Native speech+text in/out; new LFM-based INT4 QAT audio detokenizer 8× faster than LFM2-Audio's Mimi on mobile CPU. Japanese variant (Audio-1.5B-JP) in June. |
| LFM2.5-350M | 350M | Mar 31 2026; LFM2-350M continued from 10T to 28T tokens plus large-scale RL; tool-use / extraction focus. Base of the Embedding-350M and ColBERT-350M retrievers (Jun 18) and Encoder-350M (Jul 28). |
| LFM2.5-VL-450M | 450M | Apr 8 2026; 28T-token VL refresh adding grounding/bounding boxes (RefCOCO-M 81.28), function calling and 8-language image understanding. |
| LFM2.5-8B-A1B | 8.3B | May 28 2026; hybrid MoE (24 layers: 18 double-gated conv + 6 GQA, 32 experts, top-4), 38T tokens (from LFM2-8B-A1B's 12T), 128K context via 2T + 400B mid-training stages, 128K vocab, reasoning-only. AA Intelligence Index v4.3 score 7 (file's top-level anchor). DSpark drafter (328M) gives ~2.5× decode. |
| LFM2.5-230M | 230M | Jun 25 2026; smallest LFM (14 layers: 8 conv + 6 GQA), 19T tokens with 32K context extension, SFT distilled from LFM2.5-350M + DPO + multi-domain RL. 213 tok/s on a Galaxy S25 Ultra. |
| LFM2.5-Embedding-350M / ColBERT-350M | 350M | Jun 18 2026; first bidirectional LFMs (non-causal attention + convs) on LFM2.5-350M-Base, 11-language retrieval. |
| LFM2.5-Encoder-230M / 350M | 350M | Jul 28 2026; masked-LM bidirectional encoders (8K context) for classification/token tasks, with released task heads (PII detector, prompt router, spellchecker, policy linter, diffusion). |
| LFM2.5-2.6B | 2.69B | Aug 4 2026; 30 layers (22 conv + 8 GQA), 34T tokens, 128K context and vocab, pure reasoning model post-trained for agents (SFT → teacher specialization → MOPD → agentic GRPO in Hermes Agent / OpenClaw). AA Intelligence Index v4.3 score 8 — highest in family. Base checkpoint released; free on OpenRouter. |
| LFM2.5-VL-3B | 3B | Aug 12 2026; LFM2.5-2.6B backbone + SigLIP2 NaFlex 400M encoder, non-reasoning. ScreenSpot-v2 80.7, RefCOCO P@1 87.9, ToolSandbox 59.5, BLINK 61.5; 228 tok/s on M5 Max under 3.3 GB. |
| LFM2.5-DSpark drafters | 328M | Aug 20 2026; DSpark speculative-decoding draft models (296M for 1.2B-Instruct, 328M for 2.6B and 8B-A1B), up to 3.18× throughput on H100 and 2.87× on-device with identical outputs. |