Liquid AI's Qwen-Beating Claim Didn't Survive a Recount
LFM3-1.7B trended on Hugging Face after Liquid AI said its 1.73B model beats a rival five times its size. An independent recount found the lead vanishes once you stop counting one benchmark three times.

LFM3-1.7B hit the Hugging Face trending page claiming something that should stop any engineer mid-scroll: a model you can run on a phone beating a rival five times its size. The engineering is real, the headline number is not.
Liquid AI shipped LFM3-1.7B on Tuesday, Oct. 27, as the text model anchoring a six-model family—vision, audio, omni, and two decision models—all branching from one 1.73-billion-parameter backbone. The tri-hybrid architecture splits 24 layers between gated short convolutions, sliding-window attention, and global attention, trained for agentic tool-calling workloads. At 4-bit precision it fits in about 1 GB of memory. On a single Nvidia B300 it can produce 3.6 billion tokens a day at 1,024 concurrent requests.
The company also ran it against Qwen3.5-2B on 65 everyday agent tasks in real macOS environments. LFM3-1.7B won: 64% task success versus 48%, 58% fewer tokens per task, stuck less than 1% of attempts versus 15% for Qwen. That's a real, independently reproducible test.
The number that doesn't hold
Liquid AI's headline: on its own 29-benchmark table, LFM3-1.7B averages 49.37 against 47.24 for Qwen3.5-9B, a model five times its size. That 2.13-point lead is the whole case for "small models crossed a threshold."
An independent recompute found the lead rests on one evaluation counted three times across the grid. Strip the duplication and the win becomes a 0.04-point loss. Translation: the number that made this model trend is an artifact of scorecard construction, not a genuine advantage.
Liquid AI's own documentation is blunt about what it doesn't do: not recommended for agentic coding or complex multi-step reasoning. On Qwen's strengths it loses badly—80.40 to 63.37 on GPQA Diamond, 87.11 to 65.56 on RULER at 128K tokens. For general chat or broad reasoning, a larger model will outperform it by a wide margin.

The license has a ceiling too
The weights download without a gate, but "open" is doing heavy lifting. The LFM Open License v1.0 ends free commercial rights the moment your company's annual revenue exceeds $10 million USD across your whole corporate group. Past that line you need a commercial license from Liquid AI. That makes it source-available with a revenue trigger, not an OSI open-source license—useful for a solo developer, a different calculation for anything built to scale.
What to watch: whether Liquid AI corrects the benchmark table now that the duplication is public, and whether download counts (42,400 and climbing) keep rising after the correction lands or fade after the quality floor gets tested in production.
- LiquidAI/LFM3-1.7B · Hugging Face
- LFM3-1.7B - Liquid AI Docs
- LiquidAI/LFM3-1.7B-Base · Hugging Face
- Introducing LFM3: Every Modality for Specialized Intelligence
- LiquidAI LFM3-1.7B: A Local Agent Model That Fits on a Phone
- LiquidAI LFM3-1.7B: A 1GB Local AI Model Built for Agents
- Build specialized multimodal models with LFM3
- LFM3 explained: Liquid AI's 1.7B model, benchmarks, speed, license
- LFM3-1.7B Benchmarks: A 1.7B Model Built for Agentic Tool Use
- LiquidAI/LFM3-Omni-2B · Hugging Face
- LiquidAI/LFM3-Audio-2B · Hugging Face
- Add evaluation results for LiquidAI/LFM3-Audio-2B
- LiquidAI/LFM3-VL-2B · Hugging Face
- LiquidAI (Liquid AI)
- LFM2 Technical Report
- LFM3: Technical Deep Dive into Liquid AI's Tri-Hybrid Multimodal Family - Local AI Zone
- LFM License
- LFM Open License User Guide - Liquid Docs
- Liquid AI LFM3 Guide: 2B Edge Models & $10M License Cap
- LICENSE · LiquidAI/LFM2-350M-ENJP-MT at main
- Liquid AI Opens LFM3 Free to Firms Under $10M in Revenue - FourWeekMBA
- LFM2.5 from Liquid AI: above ten million dollars, the licence is not granted — CSIDIA
- LFM3's 2.13 point lead rests on one test; $10M revenue licence limit
- LiquidAI/LFM2-2.6B-Exp-GGUF · LFM Open License v1.0 isn't free license
- LiquidAI LFM2.5 Review 2026: Edge AI on a Raspberry Pi — With a License Catch