Liquid AI's 2.6B Model Beats 9B Rivals on Tool Use, Targeting the Edge Agent Market
LFM2.5-2.6B runs on consumer CPUs and phones, outperforming models 4x its size on instruction-following and tool use.
3. Liquid AI's 2.6B Model Beats 9B Rivals on Tool Use, Targeting the Edge Agent Market
Liquid AI published LFM2.5-2.6B on August 4, 2026, positioning it as a local agent model that runs in under 2.5 GB of memory. On an Apple M5 Max it delivers 220 tokens per second; on an AMD Ryzen CPU, 113 tokens per second. The model was pre-trained on roughly 34 trillion tokens, with a mid-training phase extending the context window to 128K. Post-training involved four stages: supervised fine-tuning weighted toward agentic data, domain-specific teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning run inside real agent harnesses.
The benchmark numbers are the competitive story. LFM2.5-2.6B tops every instruction-following benchmark in Liquid's evaluation set, beats both Gemma 4 variants (5.1B and 8B) on tool use, and matches Qwen3.5-9B (9.7B parameters) on ToolSandbox and Claw-Eval while running at roughly a quarter of the parameter count. That puts direct pressure on Google's Gemma small-model line and Alibaba's Qwen3.5-4B, both of which have been the default recommendations for on-device agent work. The one honest caveat Liquid flags: coding benchmarks still favor the larger models, so LFM2.5-2.6B is not a universal replacement at the edge.
The broader pattern is a compression race at the sub-3B tier. Microsoft's Phi-4-mini and Google's Gemma 4 2B have been fighting for the same deployment slot: capable enough for agentic tasks, small enough to avoid cloud inference costs. Liquid's entry, trained inside real harnesses rather than synthetic agentic data alone, is a methodological bet that environment-grounded RL generalizes better than scale. Day-one support across llama.cpp and MLX means practitioners can test that bet immediately, without waiting for ecosystem catch-up.