Innovation & Tech

Liquid AI Ships LFM3: A 1.7B Non-Transformer Model Family

Liquid AI's LFM3 family pairs a 1.7B hybrid non-transformer backbone with six models across text, vision, audio, omni and decision tasks. License terms are undisclosed.

By Olivia Hart

6 min read

Liquid AI Releases LFM3 Family, Led by 1.7B-Parameter Hybrid Model
Liquid AI Releases LFM3 Family, Led by 1.7B-Parameter Hybrid Modelstevepb / pixabay

What's News

  • Liquid AI released LFM3, a six-model family led by a 1.7B-parameter hybrid non-transformer backbone pre-trained on 22 trillion tokens.
  • At 128K-token context, LFM3-1.7B needs 1.6 GB of KV cache versus 4.3 GB for Qwen3.5-4B, per vendor figures.
  • Vendor benchmarks show LFM3-1.7B leading on agentic tasks (LiveClaw 45.6 vs 35.8 for Qwen3.5-9B) but trailing on coding and knowledge exams.
  • License terms for the LFM3 weights were not specified in the announcement.
  • LiquidCodec, a 74M-parameter audio codec, encodes speech at ~361× real time on a single CPU thread, the company says.

Liquid AI has released LFM3, a six-model production family built on a 1.7B-parameter hybrid non-transformer backbone pre-trained on 22 trillion tokens, with open weights on Hugging Face and API access through its own platform and ngrok. The announcement did not specify licensing terms for the weights — a gap that leaves commercial adoption terms undefined.

The flagship LFM3-1.7B serves as a shared backbone for six models spanning text, vision, audio, omni, and decision tasks, all in the 2B-parameter range. Liquid AI says the architecture interleaves gated short convolutions, sliding-window attention, and global attention across 24 layers.

What does the new architecture change?

The previous generation, LFM2, mixed short convolutions and attention. For LFM3, Liquid AI says it ran an automated hardware-in-the-loop architecture search across 14 device and backend combinations — including a Samsung Galaxy S26 Ultra, an OnePlus 15, a Raspberry Pi 5, an Apple M4 Max laptop, an AMD Ryzen AI laptop, Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU — to optimize for agentic workloads that need long-context reasoning at low latency.

The resulting LFM3-1.7B-Base stacks 13 gated short-convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global-attention layers. Each layer applies what the company calls "sandwich norm" — normalizing both input and output — plus head-wise gated attention, where a sigmoid gate can switch off individual heads. The vocabulary is 128K tokens, shared with LFM2.5-8B-A1B and built for non-Latin scripts.

The hybrid design enables a claimed 256K-token context window. At 128K tokens, a full context needs 1.6 GB of KV cache versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B, according to the company.

How do the six models differ?

All six branch from LFM3-1.7B-Base:

  • LFM3-1.7B — text model for local agents, tool calling, structured output, and domain fine-tuning.
  • LFM3-VL-2B — vision-language model with video input and spatial and embodied 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 pipelines differ per model. The text model goes through supervised fine-tuning on 1.75T tokens, mixed reinforcement learning, multi-domain on-policy distillation from six specialist RL teachers, and agentic RL in real harnesses covering web browsing and office work. The omni model merges text, vision, and audio RL stages.

Do the benchmarks hold up?

Every number in Liquid AI's benchmark tables was 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, ToolSandbox at 80.4 versus 76.4, and IFEval at 96.7 versus 92.1.

The picture shifts on harder suites. Qwen3.5-9B edges ahead on Claw-Eval pass³ (51.9 versus 51.4) and EnterpriseOps-Gym (28.97 versus 28.20). Larger models win on coding — SWE-bench Verified at 43.4 versus 51.2 — and on knowledge exams: GPQA Diamond at 63.4 versus 80.4, and Humanity's Last Exam with tools at 13.9 versus 17.5. The company itself says 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; it scored 49.2 overall average in the ≤2B active-parameter group, behind Qwen3-1.7B at 53.8 and granite-4.0-h-tiny at 53.9, though ahead of several larger models. Liquid AI's own reported suite average is 49.37 — closely matching the independent figure.

What about speed on real hardware?

Liquid AI reports edge inference for Q4₀ GGUF in llama.cpp at 32K context. On an AMD Ryzen AI Max+ 395, the model hits 3,771 tok/s prefill and 129.6 tok/s decode at 2.75 GB of memory. On an Apple M5 Max, it reaches 8,028 tok/s prefill and 237.1 tok/s decode at 2.25 GB. On an iPhone 17 Pro, it delivers 1,061 tok/s prefill and 43.6 tok/s decode at 1.90 GB. In each case, LFM3-1.7B posts the highest prefill, highest decode, and lowest memory among compared models.

On a single NVIDIA B300 in BF16 with SGLang, the company reports 3.6B tokens per day at 1,024 concurrent requests, with p95 latency of 86–158 ms at 10–600 req/s for 128-in/64-out workloads, saturating at 1,000 req/s. Competing models on vLLM saturate at far lower request rates, the company says. All these figures are vendor-measured on specific hardware and custom software builds — independent replication has not followed.

Why does the audio codec matter?

LFM3-Audio-2B and LFM3-Omni-2B use LiquidCodec, a 74M-parameter streaming audio codec that replaces the Mimi codec used in earlier models. It moves from residual vector quantization to grouped finite scalar quantization, enabling parallel token prediction via MaskGIT-style refinement.

Liquid AI reports encoding at roughly 361× real time and decoding at roughly 87× real time on a single Apple M4 Max CPU thread in FP32. Reconstruction quality matches or exceeds Mimi, X-Codec 2, and NanoCodec across eight test sets, and zero-shot TTS naturalness leads on all eight. One caveat: token incompatibility means existing speech models cannot swap in LiquidCodec without retraining.

What remains open?

Weights are available on Hugging Face in BF16, FP8, NVFP4, GGUF, MLX, ONNX Runtime, and OpenVINO variants, and ngrok's AI Gateway lists liquid/lfm3-1.7b, the VL, audio, and omni models, plus liquid/d1, liquid/d1.1, and liquid/d1-ultra. But license terms for commercial use, redistribution, and fine-tuning remain unspecified. So do independent replications of the agentic benchmarks, independent stress tests of the claimed 256K context window, and the training data composition beyond "22T tokens" plus a 3.5T-token mid-training stage on curated agentic traces.

The release is a rare production-scale deployment of a non-transformer architecture, and the memory and throughput numbers suggest real gains for on-device agents. Until the license terms land and third parties replicate the benchmarks, developers should treat the efficiency claims as promising but unverified.

Source: epyhia.com

Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,203 articles

Related articles

« Previous articleNext article »