LFM3-1.7B

LFM3-1.7B

Quick answer: LFM3-1.7B is Liquid AI's 1.73B-parameter text model built on a new "tri-hybrid" architecture (gated short convolutions + sliding-window attention + global attention), released October 2026 with a 256K-token context window. Post-trained for agentic harnesses (Hermes Agent, Pi, OpenClaw), it scores 71.4% on BFCL V4, 80.4% on ToolSandbox, and 96.7% on IFEval — Liquid reports it can power agentic applications that previously needed ~9B-parameter models. At 4-bit precision it needs about 1GB of memory, fitting on a phone or laptop.

At a Glance

Where LFM3-1.7B leads

• 71.4% BFCL V4 and 80.4% ToolSandbox — ahead of every compared model up to 9B parameters (LFM2.5-2.6B, MiniCPM5-2B, Gemma 4 E4B, Qwen3.5-4B, and even Qwen3.5-9B on BFCL V4)
• 96.7% IFEval, 73.0% IFBench, 89.1% Multi-IF — leads all compared models on instruction-following
• 45.6 LiveClaw and 41.3 AutomationBench (public) — large agentic-task gains over its LFM2.5-2.6B predecessor (14.7 / 0.9)
• ~1GB memory at 4-bit quantization; 2.5x+ faster prefill and 1.8x+ faster decode than compared models across AMD/Apple/iPhone hardware tested

Where it lags

• 63.4% GPQA Diamond — trails Qwen3.5-4B (75.8%) and Qwen3.5-9B (80.4%) on graduate-level science knowledge
• 7.2% HLE (no tools) — behind Qwen3.5-9B (16.4%); not recommended by Liquid AI for knowledge-heavy tasks
• Not recommended by Liquid AI for agentic coding or complex multi-step reasoning given its small size

Best for: local/edge agent deployments — tool calling, structured outputs, and domain-specific fine-tuning — where on-device privacy and near-zero marginal inference cost matter more than frontier coding or deep knowledge recall.

What LFM3-1.7B Is

LFM3-1.7B is Liquid AI's newest small model for local agents, introducing a "tri-hybrid" architecture: 24 layers combining 13 gated short-convolution blocks, 5 sliding-window attention blocks, and 6 global attention blocks. It was pre-trained on 22T tokens across two stages (a general mix plus a mid-training phase adding scientific papers, repository-level code, math, and agentic/tool-call traces while extending context to 256K tokens), then post-trained through supervised fine-tuning, mixed reinforcement learning, multi-domain on-policy distillation from specialist RL teachers, and a final agentic-RL stage run inside real harnesses.

Liquid positions LFM3-1.7B explicitly as a successor that "can power agentic applications that used to need 9B parameters" — its own benchmark table shows it beating same-class and larger models (Qwen3.5-4B, Qwen3.5-9B) on most agentic, tool-use, and instruction-following evaluations while trailing on knowledge-heavy and long-context-retrieval tasks (GPQA Diamond, HLE, RULER). The model also supports "deferred tools" (tool search), letting a harness keep large tool catalogs out of the prompt until needed — a capability aimed squarely at production agent deployments with many available tools.

Specifications

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face in BF16, FP8, NVFP4, GGUF (llama.cpp/Ollama/LM Studio), MLX (Apple Silicon), and ONNX formats. At 4-bit quantization, the model uses about 1GB of memory; at 1,024 concurrent requests it can produce 3.6 billion tokens/day on a single NVIDIA B300.

Public Benchmark Scores

Scores above are reported by Liquid AI and shown for context. They are not Benchgen measurements. Figures depend on harness, tools, and effort settings — all models in Liquid's comparison table were served with vLLM in BF16 at full context length; see the source for full methodology.

Excluded from the table above (bucket-C, documented per skill procedure):

• LiveClaw (45.60) — Liquid's own OpenClaw-harness agent eval; no independently discoverable public benchmark definition/paper found (same exclusion pattern as LFM2.5-2.6B's prior release).
• BrowseComp-Plus (OpenClaw subset) (37.00) — a harness-specific subset of BrowseComp-Plus; our local browsecomp* entities don't document this exact subset's methodology, so left unmapped rather than guessed.
• WebArena-Verified (38.42) and EnterpriseOps-Gym (28.20) — no local Benchgen benchmark entity exists for either; insufficient public methodology/paper info in the model card alone to stand up new configs this run.
• τ³-bench (4 domains), average (68.01) — ambiguous generic name vs. our existing domain-specific tau3-banking entity; per standing policy, never silently narrow/widen scope. Excluded, same as the Nanbeige4.3-33B precedent.
• IFStruct (93.72) — Liquid's own structured-instruction-following eval; not an independently verifiable standalone public benchmark (same exclusion as LFM2.5-2.6B).
• Humanity's Last Exam (with tools) (13.90) and AA-Omniscience accuracy/non-hallucination sub-metrics (7.07 / 92.93) — secondary/derived metrics alongside the primary HLE and AA-Omniscience scores already listed above; not separate benchmark entities.
• RULER 128K (65.56) — no local Benchgen entity for RULER exists yet; skipped rather than creating a new config this run (candidate for a future bucket-B review given it's a widely cross-model-comparable long-context benchmark).

LFM3-1.7B vs Alternatives

Despite being Liquid AI's smallest LFM3-generation model, LFM3-1.7B leads every compared model — including the 9.7B Qwen3.5-9B — on BFCL V4, ToolSandbox, and IFEval, while remaining small enough to run locally on a phone or laptop; the clear trade-off for builders is weaker raw knowledge recall (GPQA Diamond, HLE) versus the larger Qwen3.5 models.

How LFM3-1.7B Performs on Real Agent Tasks

LFM3-1.7B's benchmark spread is heavily weighted toward tool calling, instruction following, and short-horizon agentic tasks (BFCL V4, ToolSandbox, MCP-Atlas, Claw-Eval) rather than deep knowledge or long-horizon coding — consistent with Liquid's explicit guidance that the model is "not recommended for agentic coding or complex multi-step reasoning due to its small size." A high public score on tool-use benchmarks indicates strong general tool-calling reliability, but builders deploying it for a specific production agent (e.g. a fixed set of internal APIs) should still validate accuracy on their own tool schemas and task distribution rather than relying on the aggregate benchmark numbers alone.

Use LFM3-1.7B via API

from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="LiquidAI/LFM3-1.7B",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "What is C. elegans?"}]
output = generator(
    messages,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.1,
    max_new_tokens=2048,
)
print(output[0]["generated_text"][-1]["content"])

Related Models & Benchmarks

• Compare: LFM2.5-2.6B · Qwen3.5-9B
• Benchmarks: BFCL V4 · ToolSandbox · GPQA Diamond

Frequently Asked Questions

Specs and scores sourced from Liquid AI's official Hugging Face model card and release blog post (liquid.ai/blog/lfm3); third-party/comparison-model scores (LFM2.5-2.6B, MiniCPM5-2B, Gemma 4 E4B, Qwen3.5-4B/9B) as reported in the same launch benchmark table. Last updated 2026-10-29.