Liquid AI's LFM2.5 Triples Training Data to 38T Tokens and Cuts Hallucinations by 9x on Device
Liquid AI on May 28 released LFM2.5-8B-A1B, the follow-up to its October 2025 edge MoE model, with three significant changes: context window up from 32K to 128K tokens, pretraining data tripled from 12T to 38T tokens, and a full shift to reasoning-only output via chain-of-thought before every answer. The result is a model that runs on a laptop and now competes on benchmarks where it previously scored near zero.
By the Numbers
The comparison against the prior LFM2-8B-A1B is stark:
| Benchmark | LFM2-8B-A1B | LFM2.5-8B-A1B | Delta |
|---|---|---|---|
| MATH500 | 74.80% | 88.76% | +13.96pp |
| AIME25 | 20.00% | 42.53% | +22.53pp |
| IFEval (instruction following) | 79.44% | 91.84% | +12.40pp |
| BFCLv3 (tool calling) | 45.07% | 64.36% | +19.29pp |
| BFCLv4 (tool calling) | 25.52% | 48.50% | +22.98pp |
| Tau² Telecom (agentic) | 13.60% | 88.07% | +74.47pp |
| Tau² Retail (agentic) | 7.02% | 39.82% | +32.80pp |
| Non-hallucination rate (AA-Omniscience) | 7.46% | 63.47% | +56.01pp |
The Tau² Telecom jump from 13.6% to 88.1% is the most striking. It places LFM2.5 inside the competent-agent range for that benchmark — not at the frontier (Claude Opus 4.7 exceeds 99% telecom), but within striking distance of mid-tier models running at orders-of-magnitude more active compute.
Why Reasoning-Only Changes the Edge Math
The model’s move to reasoning-only is an architectural bet: MoE models with few active parameters run in compute-bound settings, where each extra reasoning token is cheap relative to what it buys. Liquid AI’s argument is that the usual flash-inference tradeoff (skip reasoning, go fast) doesn’t apply when you’re bottlenecked on memory bandwidth rather than flops. If each token is cheap to generate, reasoning tokens are nearly free — so you might as well use them.
That also affects hallucinations. The old model had a 7.5% non-hallucination rate, meaning it confidently answered wrong nearly all the time. The new model hits 63.5%. Liquid AI added a targeted RL stage using an avg@k-based reward over a diverse knowledge dataset, explicitly reinforcing abstention on queries outside the model’s reliable knowledge. The goal is a sharper knowledge boundary — the model should say nothing rather than confabulate.
Doom Loops and Vocabulary Expansion
Two other changes worth noting:
Doom loops. Long reasoning traces in small models tend to get stuck in loops. Liquid AI added a preference optimisation stage that redistributes probability mass away from loop-inducing tokens in context, and a lightweight shaping reward during RL that penalises excessive use of restart words (“Wait…”, etc.). It’s a targeted fix for a specific failure mode that gets worse at small scale.
Vocabulary doubling. The tokeniser expanded from 65K to 128K using in-place extension rather than retraining from scratch — keeping existing token IDs unchanged and initialising new embedding rows as the mean of their sub-token decompositions. The efficiency gain for non-Latin scripts is material: Hindi +120%, Thai +238%, Vietnamese +118%. Arabic +38%.
Availability and Stack
LFM2.5-8B-A1B ships with day-one support for llama.cpp, MLX, vLLM, and SGLang. Base and post-trained weights are on Hugging Face. It runs on entry-level consumer hardware.
The architecture is the same MoE, GQA, and gated short convolution combination as the prior model — no structural novelty, just better training. Liquid AI says it’s the fastest in its size class on both CPU and GPU inference.