Liquid AI introduced LFM3 on Tuesday: an open-weight family in the 2-billion-parameter range that branches six models off a single base backbone — text, vision, audio, an omni model and two decision models. The pitch is not that these beat the biggest models. It is that a small model you can fine-tune on one GPU is better at your specific job than a generalist you rent through an API.
The architecture
LFM3-1.7B-Base has 24 layers in three flavours: 13 gated short convolutions with a kernel size of 7, five sliding-window attention layers with a 256-token window, and six layers of global attention. Liquid calls it a tri-hybrid mix, produced by an architecture search that ran hardware-in-the-loop across 14 device and backend combinations — phones, a Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs and Qualcomm's Hexagon NPU.
Several details point at the same goal. Head-wise gated attention lets a sigmoid gate scale down each head's contribution when it has nothing to add, which also suppresses activation outliers that hurt quantization. Global attention uses partial RoPE; each block normalizes both its input and its output. The vocabulary is 128,000 tokens, tuned to tokenize non-Latin scripts more efficiently.
Pre-training covered 22 trillion tokens. The final 3.5 trillion were a mid-training stage that extends the context window to 256,000 tokens and raises the share of agentic traces and tool calls. Four post-training stages followed: supervised fine-tuning, mixed reinforcement learning with a length penalty, on-policy distillation from six specialist RL teachers, and agentic RL run inside real agent harnesses.
What the numbers say — and what they don't
On its own average across the reported benchmarks, LFM3-1.7B scores 49.37, ahead of Qwen3.5-9B (47.24) and far ahead of Gemma 4 E4B (31.74). It leads on tool use and instruction following: BFCL V4 71.39, ToolSandbox 80.36, IFEval 96.66. It loses clearly on agentic coding — SWE-bench Verified 43.40 against 51.20 for the roughly six-times-larger Qwen3.5-9B — and on world knowledge. That is not a flaw in the announcement; it is its most honest part. Size still decides where knowledge and long repair loops matter.
One result stands out anyway: on AA-Omniscience the non-hallucination rate is 92.93 percent, against 52.59 percent for Qwen3.5-9B. A small model that more often admits it does not know is frequently more useful for decision work than one that guesses confidently.
On the device
With llama.cpp at Q4_0 and a 32K context, Liquid reports 1.8 to 2.8 times faster decoding than the next-best comparison model, at the lowest memory footprint. A full 128K context costs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. On a single NVIDIA B300 with SGLang in BF16 the company claims 3.6 billion tokens per day at 1,024 concurrent requests; to demonstrate the throughput it ran more than 300,000 agents in parallel over an email archive.
The vision model takes video for the first time and averages 71.4 across 62 vision benchmarks — level with Gemma 4 12B, a model six times larger, per Liquid's table. The audio model covers 18 languages with an in-house encoder and codec and turns text into speech at about 87 times real time on a single CPU thread. The two decision models, d1-2B and d1-Omni-2B, answer rather than write: many questions in one forward pass, each with a calibrated probability, including bounding boxes on an image and exact text spans from a document.
The real test
Two fine-tuning examples carry the argument. A vision model trained for a few hours on one A100 with about 3,000 photos of wind-turbine blades lifts classification accuracy from under 10 percent to 67 percent. A robot controller built as a rank-32 LoRA adapter in six GPU-hours on a single AMD MI325X reaches 96.2 percent average success across the four LIBERO suites.
Whether that holds up is not decided by a blog post but by the first user who puts a 2B model on their own data and finds it sufficient. The benchmarks are vendor-run; independent reproductions are still missing. Anyone planning around LFM3 today should check two things: whether the task is genuinely narrow, and whether the error rate in their own operation is lower than the generalist's.
