LFM3-1.7B vs LFM2.5-2.6B: smaller, longer context, stronger on agents
LFM3-1.7B has 35% fewer parameters than LFM2.5-2.6B, twice the context (256K vs 128K) and averages 49.4 vs 31.6 on Liquid AI’s benchmark suite. Specs, scores and API prices compared.
30 окт. 2026 г. · API ManyLFM2.5-2.6B was Liquid AI’s strongest small text model before LFM3. Liquid says the new LFM3-1.7B outperforms it while using 35% fewer parameters, thanks to a new tri-hybrid architecture and training recipe.
Source:Liquid AI: Introducing LFM3 (Oct 27, 2026)
Specs and prices
| LFM3-1.7B | LFM2.5-2.6B | |
|---|---|---|
| Parameters | 1.73B | 2.69B |
| Layers | 24: 13 gated short convolutions, 5 sliding-window attention, 6 global attention | 30: 22 double-gated short convolutions, 8 grouped-query attention |
| Pre-training tokens | 22T | 34T |
| Context window | 256K tokens | 131,072 tokens |
| Vocabulary | 128K | 128K |
| Languages | Same 16 languages | Same 16 languages |
| License | LFM Open License v1.0 | LFM Open License v1.0 |
| API input / output per 1M tokens | $0.08 / $0.20 (Liquid API) | $0.35 / $0.50 (OpenRouter, Liquid provider) |
| Cached input per 1M tokens | $0.02 | $0.04 |
| On API Many | No | No |
Sources:Hugging Face: LiquidAI/LFM3-1.7B model card · Hugging Face: LiquidAI/LFM2.5-2.6B model card · Liquid AI: Pricing · OpenRouter: LiquidAI LFM2.5-2.6B
Benchmarks
| Benchmark (Liquid AI’s evaluation) | LFM3-1.7B | LFM2.5-2.6B |
|---|---|---|
| LiveClaw (agents) | 45.60 | 14.70 |
| Claw-Eval General (EN), average | 68.64 | 62.63 |
| WebArena-Verified | 38.42 | 3.02 |
| AutomationBench (public) | 41.29 | 0.94 |
| BFCL V4 (function calling) | 71.39 | 55.89 |
| ToolSandbox | 80.36 | 78.04 |
| SWE-bench Verified (OpenCode) | 43.40 | 18.00 |
| Terminal-Bench 2.1 (OpenCode) | 12.92 | 0.00 |
| LiveCodeBench v6 | 73.46 | 60.57 |
| IFEval | 96.66 | 95.39 |
| IFBench | 73.03 | 60.48 |
| GPQA Diamond | 63.37 | 57.78 |
| AA-Omniscience accuracy (public) | 7.07 | 8.75 |
| AA-Omniscience non-hallucination (public) | 92.93 | 53.69 |
| AA-LCR (long context) | 52.83 | 6.00 |
| RULER 128K | 65.56 | 1.15 |
| Average of all 29 scores | 49.37 | 31.55 |
What changed
- Architecture: LFM3 keeps short convolutions for over half the layers but adds sliding-window attention next to global attention, uses head-wise gated attention and partial RoPE, and widens the convolution kernel from 3 to 7.
- Long context: the context window doubles to 256K, and the long-context scores move from near zero (RULER 128K 1.15) to usable (65.56).
- Agents: the largest gains are in browser and business-app agents (WebArena-Verified, AutomationBench) and agentic coding, after a new agentic RL stage run inside real agent harnesses.
- Size: 35% fewer parameters; Liquid says the 4-bit text model fits in about 1 GB.
Sources:Liquid AI: Introducing LFM3 (Oct 27, 2026) · Hugging Face: LiquidAI/LFM3-1.7B model card
Which one should you use?
- New projects: LFM3-1.7B scores higher on almost every benchmark Liquid published, is smaller, and is cheaper on Liquid’s API than LFM2.5-2.6B is on OpenRouter.
- If you already call LFM2.5-2.6B through OpenRouter: LFM3-1.7B is not on OpenRouter yet, so switching today means Liquid’s own API or self-hosting.
- If you have LFM2.5 fine-tunes: LFM3 is a new architecture, so adapters do not carry over; plan a new fine-tuning run on the LFM3-1.7B base.
- Both models: Liquid does not recommend them for complex multi-step reasoning or knowledge-heavy work without retrieval.
Sources:Hugging Face: LiquidAI/LFM3-1.7B model card · Hugging Face: LiquidAI/LFM2.5-2.6B model card · OpenRouter: Liquid models
Need a small hosted model on API Many today? GPT-6 Luna starts at $0.10 вход · $0.50 выход and Gemini 3.5 Flash Lite at $0.30 вход · $2.499 выход, both through one OpenAI-compatible endpoint.
Related
- LFM3: what Liquid AI released, how it scores, and where to run itThe full LFM3 family, including LFM3-Omni-2B.
- Spark-X2.6-4B: the 4B open-weight Spark model, explainedAnother small open-weight model built for on-device agents.
FAQ
Is LFM3-1.7B better than LFM2.5-2.6B?
On Liquid AI’s published benchmarks, yes: it averages 49.37 versus 31.55 across 29 scores and leads on agents, tool use, coding, instruction following and long context, despite having 35% fewer parameters.
Which has the longer context window?
LFM3-1.7B: 256K tokens versus 131,072 for LFM2.5-2.6B.
Which is cheaper to call?
LFM3-1.7B costs $0.08 / $0.20 per 1M input / output tokens on Liquid’s API. LFM2.5-2.6B costs $0.35 / $0.50 on OpenRouter, which also lists a free variant with a 65,536-token context.
Can I use either model on API Many?
Not currently. Neither LFM3-1.7B nor LFM2.5-2.6B is offered on API Many.
Sources
- Hugging Face: LiquidAI/LFM3-1.7B model card — LFM3-1.7B specs and the benchmark table with LFM2.5-2.6B.
- Liquid AI: Introducing LFM3 (Oct 27, 2026) — Architecture changes, the 35% parameter claim and training stages.
- Hugging Face: LiquidAI/LFM2.5-2.6B model card — LFM2.5-2.6B specs.
- Liquid AI: Pricing — LFM3-1.7B API price.
- OpenRouter: LiquidAI LFM2.5-2.6B — LFM2.5-2.6B price, context and output limit on OpenRouter.
- OpenRouter: Liquid models — Liquid models listed on OpenRouter (no LFM3 as of Oct 30, 2026).
Все модели из статьи доступны в API Many. Оплата по факту от $10.