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

Liquid AI has released the LFM3 family of models—its latest production lineup built on an evolved hybrid non-transformer architecture. 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.

The company announced the family on its blog and published open weights on Hugging Face, with API access through its own platform and partners including ngrok. Liquid AI says LFM3-1.7B was pre-trained on 22 trillion tokens and uses a hybrid architecture that interleaves gated short convolutions, sliding-window attention, and global attention across 24 layers [5].

A non-transformer backbone for on-device deployment

Liquid AI's previous generation, LFM2, used a hybrid of short convolutions and attention. For LFM3, the company says it ran an automated hardware-in-the-loop architecture search across 14 device and backend combinations—phones (Samsung Galaxy S26 Ultra and OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max and AMD Ryzen AI), Arm and x86 cloud CPUs, plus GPUs and the Qualcomm Hexagon NPU—to optimize for agentic workloads that need long-context reasoning at low latency [5].

The resulting LFM3-1.7B-Base stacks 13 gated short-convolution layers, 5 sliding-window attention layers (256-token window), and 6 global-attention layers. Each layer applies "sandwich norm" (normalizing both input and output) and head-wise gated attention, where a sigmoid gate can switch off individual heads. Global attention uses partial rotary position embeddings (only the highest-frequency half of each head), while sliding-window layers keep full RoPE. The vocabulary is 128K tokens, shared with LFM2.5-8B-A1B, built for non-Latin scripts [5].

The company says the hybrid design enables a 256K-token context window. At 128K-token context, 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 [5].

Six models from one backbone

All six LFM3 models branch from LFM3-1.7B-Base:

• LFM3-1.7B – text model for local agents, tool calling, structured output, domain fine-tuning.
• LFM3-VL-2B – vision-language model with video input, step-by-step reasoning, spatial/embodied reasoning.
• LFM3-Audio-2B – audio model (text/audio in, text/speech/both out) with in-house audio encoder and LiquidCodec tokenizer, 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 [5].

Post-training pipelines differ per model. The text model goes through SFT (1.75T tokens), mixed RL, multi-domain on-policy distillation (MOPD) from six specialist RL teachers, and agentic RL in real harnesses (web browsing, office work). The vision and audio models add modality-specific mid-training and RL stages. The omni model merges text, vision, and audio RL [5].

Benchmarks: vendor-run, agentic focus

Liquid AI publishes extensive benchmark tables comparing LFM3-1.7B against MiniCPM5-2B, Gemma 4 E4B, Qwen3.5-4B, Qwen3.5-9B, and its own LFM2.5-2.6B. All evaluations were run by Liquid AI using vLLM in BF16 at full context length [5][6].

On agentic tasks, LFM3-1.7B leads the compared set on LiveClaw (45.6 vs. 35.8 for Qwen3.5-9B), WebArena-Verified (38.4 vs. 28.9), AutomationBench (41.3 vs. 15.1), BFCL V4 (71.4 vs. 62.8), ToolSandbox (80.4 vs. 76.4), and IFEval (96.7 vs. 92.1). On Claw-Eval pass³, Qwen3.5-9B edges ahead (51.9 vs. 51.4). On EnterpriseOps-Gym, Qwen3.5-9B scores slightly higher (28.97 vs. 28.20). Larger models still win on coding benchmarks (SWE-bench Verified: 43.4 vs. 51.2 for Qwen3.5-9B) and knowledge-heavy exams (GPQA Diamond: 63.4 vs. 80.4; Humanity's Last Exam with tools: 13.9 vs. 17.5) [5][6].

The company notes that search nearly doubles its score on Humanity's Last Exam (7.17 → 13.90) and that the model is "not recommended for agentic coding or complex multi-step reasoning due to its small size" [6].

An independent evaluation in the Lara Mini base paper (arXiv:2611.16956) tested LFM3-1.7B among 56 open base models. In that evaluation, LFM3-1.7B (which was trained on 22T tokens) scored 49.2 overall average 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 [3]. Liquid AI's own reported average across its benchmark suite is 49.37 [5][6].

Edge and server inference claims

Liquid AI provides edge inference numbers for Q4₀ GGUF in llama.cpp (32K context, batch 1, 100 output tokens, GPU offload) on three devices:

• AMD Ryzen AI Max+ 395: 3,771 tok/s prefill, 129.6 tok/s decode, 2.75 GB memory.
• Apple M5 Max: 8,028 tok/s prefill, 237.1 tok/s decode, 2.25 GB memory.
• iPhone 17 Pro: 1,061 tok/s prefill, 43.6 tok/s decode, 1.90 GB memory.

In each case, LFM3-1.7B shows the highest prefill, highest decode, and lowest memory among the compared models (MiniCPM5-2B, Qwen3.5-4B, Qwen3.5-9B, Gemma 4 E4B) [5].

On a single NVIDIA B300 (BF16, SGLang), the company reports 3.6B tokens/day at 1,024 concurrent requests, and 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 [5].

These figures are vendor-measured on specific hardware and software stacks (llama.cpp builds with LFM3 support, SGLang with LFM3 support).

Audio codec as a differentiator

LFM3-Audio-2B and LFM3-Omni-2B use LiquidCodec, a 74M-parameter streaming audio codec built on the LFM3 architecture. Liquid AI says it replaces the Mimi codec used in earlier models, moving from residual vector quantization (RVQ) to grouped finite scalar quantization (GFSQ), which enables parallel token prediction via MaskGIT-style refinement [1][4].

The company reports encoding at ~361× real time and decoding at ~87× real time on a single CPU thread (Apple M4 Max, FP32), with reconstruction quality (ViSQOL, mel distance) matching or exceeding Mimi, X-Codec 2, and NanoCodec across eight test sets (LibriSpeech, Seed-TTS, FLEURS). Zero-shot TTS naturalness (UTMOSv2) leads on all eight test sets, while intelligibility (WER/CER) leads on four of the eight. Quantizing the decoder to Q4₀ shrinks it 3.5× with minimal quality loss [1][4].

LiquidCodec is available on Hugging Face and loads via Transformers. The LFM3-Audio and LFM3-Omni GGUFs include Q4₀, Q4_K_M, and Q8₀ builds for llama.cpp [1].

Availability and licensing

Weights are open on Hugging Face (LiquidAI/LFM3-1.7B and siblings) with checkpoint variants for vLLM, SGLang, Transformers (BF16, FP8, NVFP4), llama.cpp/Ollama/LM Studio (GGUF), MLX, ONNX Runtime, and OpenVINO [6]. The Liquid AI API and Playground serve the models directly. ngrok's AI Gateway also lists liquid/lfm3-1.7b, liquid/lfm3-vl-2b, liquid/lfm3-audio-2b, liquid/lfm3-omni-2b, and decision models liquid/d1, liquid/d1.1, liquid/d1-ultra [2].

The announcement did not specify licensing terms for the LFM3 weights.

What remains unknown

• License terms for commercial use, redistribution, and fine-tuning.
• Independent benchmark replication on the published agentic suites (LiveClaw, Claw-Eval, WebArena-Verified, etc.) and standard academic benchmarks (MMLU, GSM8K, HumanEval).
• Long-context quality beyond the company's RULER/ONERULER/HELMET numbers; the 256K window is claimed but not independently stress-tested.
• Audio and omni model evaluation outside Liquid's controlled codec-ablation setup; token incompatibility means existing speech models cannot swap in LiquidCodec without retraining [4].
• Training data composition beyond "22T tokens," the company's description that the mix "adds more scientific papers, repository-level code, a wider range of knowledge sources, and considerably more math," and a 3.5T mid-training stage on curated agentic traces [5].

Liquid AI's release marks a rare deployment of a non-transformer architecture at production scale. The benchmarks suggest genuine advances in on-device throughput and memory efficiency for agentic workloads, but the vendor-run evaluations, unspecified license, and limited independent audit mean the practical trade-offs for developers remain to be verified.