Liquid AI Ships LFM3: 1.7B Non-Transformer Model Runs in 1.9 GB
Liquid AI released LFM3, a 1.7B-parameter hybrid non-transformer family with a 256K context window, 43.6 tok/s decode on iPhone 17 Pro, and open weights on Hugging Face.
By Grace Kim
5 min read

What's News
- LFM3-1.7B was pre-trained on 22 trillion tokens and supports a 256K-token context window.
- At 128K context, LFM3 needs 1.6 GB of KV cache versus 5.6 GB for MiniCPM5-2B.
- On iPhone 17 Pro, LFM3-1.7B decodes at 43.6 tok/s using 1.90 GB of memory.
- The six-model family covers text, vision, audio, omni, and decision tasks, all near 2B parameters.
- License terms for the LFM3 weights were not disclosed in the announcement.
Liquid AI has released LFM3, a family of six production models led by a 1.7-billion-parameter hybrid backbone that needs just 1.6 GB of KV cache at 128K-token context — versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. The company published open weights on Hugging Face and offers API access through its own platform and partners including ngrok.
The flagship LFM3-1.7B serves as the shared backbone for six models spanning text, vision, audio, omni, and decision tasks, all in the 2B-parameter range. Liquid AI says the base model was pre-trained on 22 trillion tokens.
How does the architecture work?
LFM3 is not a transformer. The previous generation, LFM2, mixed short convolutions with attention. For LFM3, Liquid AI ran an automated hardware-in-the-loop architecture search across 14 device and backend combinations — including a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.
The resulting LFM3-1.7B-Base stacks 24 layers: 13 gated short-convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global-attention layers. Each layer uses "sandwich norm" on both input and output, plus head-wise gated attention, where a sigmoid gate can switch off individual heads. Global attention applies partial rotary position embeddings on the highest-frequency half of each head; sliding-window layers keep full RoPE. The 128K-token vocabulary is shared with LFM2.5-8B-A1B and built for non-Latin scripts. The design enables a claimed 256K-token context window.
The takeaway: this is a model architected for phones and edge devices, not data-center brute force.
What's in the six-model family?
All six models branch from LFM3-1.7B-Base:
- LFM3-1.7B — text model for local agents, tool calling, structured output, and fine-tuning
- LFM3-VL-2B — vision-language model with video input and spatial reasoning
- LFM3-Audio-2B — audio model with an in-house encoder and LiquidCodec tokenizer, covering 18 languages
- LFM3-Omni-2B — unified model taking speech, images, video, and text; outputting text and speech
- d1-2B and d1-Omni-2B — decision models specialized from the vision and omni paths
Post-training differs per model. The text model runs SFT on 1.75T tokens, mixed RL, multi-domain on-policy distillation from six specialist RL teachers, and agentic RL in real harnesses such as web browsing and office work. Vision and audio models add modality-specific mid-training and RL stages.
What do the benchmarks show?
All evaluations were run by Liquid AI itself using vLLM in BF16 at full context length. On agentic tasks, LFM3-1.7B leads the compared set: LiveClaw at 45.6 versus 35.8 for Qwen3.5-9B, WebArena-Verified at 38.4 versus 28.9, AutomationBench at 41.3 versus 15.1, BFCL V4 at 71.4 versus 62.8, and ToolSandbox at 80.4 versus 76.4.
Bigger models still win where raw capacity matters: SWE-bench Verified goes to Qwen3.5-9B, 51.2 versus 43.4; GPQA Diamond, 80.4 versus 63.4. Liquid AI states plainly that the model is "not recommended for agentic coding or complex multi-step reasoning due to its small size."
One independent data point exists. The Lara Mini base paper (arXiv:2611.16956) tested LFM3-1.7B among 56 open base models, scoring it 49.2 in the ≤2B active-parameter group — behind Qwen3-1.7B (53.8) and granite-4.0-h-tiny (53.9), but ahead of several larger models. Liquid AI's own suite average is 49.37.
How fast is it on real hardware?
Vendor-measured edge numbers, Q4₀ GGUF in llama.cpp at 32K context, batch 1:
- Apple M5 Max: 8,028 tok/s prefill, 237.1 tok/s decode, 2.25 GB memory
- AMD Ryzen AI Max+ 395: 3,771 tok/s prefill, 129.6 tok/s decode, 2.75 GB memory
- iPhone 17 Pro: 1,061 tok/s prefill, 43.6 tok/s decode, 1.90 GB memory
On a single NVIDIA B300 in BF16 with SGLang, Liquid AI reports 3.6B tokens per day at 1,024 concurrent requests, p95 latency of 86–158 ms across 10–600 req/s, and saturation at 1,000 req/s. Competing models on vLLM saturate at far lower rates, the company says.
What about the audio codec?
LiquidCodec, a 74M-parameter streaming codec, replaces the Mimi codec used earlier. It switches from residual vector quantization to grouped finite scalar quantization, enabling parallel token prediction. Liquid AI reports encoding at ~361× real time and decoding at ~87× real time on a single Apple M4 Max CPU thread, with zero-shot TTS naturalness leading on all eight test sets. A Q4₀-quantized decoder shrinks 3.5× with minimal quality loss.
Weights are available in formats spanning vLLM, SGLang, Transformers (BF16, FP8, NVFP4), GGUF, MLX, ONNX Runtime, and OpenVINO. ngrok's AI Gateway lists liquid/lfm3-1.7b and its siblings, plus decision models liquid/d1, liquid/d1.1, and liquid/d1-ultra.
What's missing: license terms for commercial use, independent replication of the agentic benchmarks, stress tests of the 256K context window, and training-data detail beyond "22T tokens" plus a 3.5T mid-training stage. Until those gaps close, developers weighing LFM3 for production edge deployment are taking Liquid AI's numbers largely on faith.
Source: epyhia.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,174 articles