LFM3-1.7B pricing, plans and limits

Pick LFM3-1.7B if you want a tool-calling model that fits in about 1 GB at 4-bit, runs on a laptop or phone and can be fine-tuned on a single GPU. Skip it for agentic coding: Liquid advises against it, and its 43.40 on SWE-bench Verified trails the 9.7B Qwen model's 51.20 in Liquid's own table.

LFM3-1.7B is a 1.73-billion-parameter open-weight text model from Liquid AI with a 256K-token context window and native tool calling. Its 24 layers mix gated short convolutions with sliding-window and global attention, and Liquid post-trained it for local agent harnesses such as Hermes Agent and OpenClaw.

Provider: Liquid AI · Family: LFM3

More about Liquid AI on Hokai

Context window: 262,144 tokens

Input modalities: text, tool-calls · Output: text, tool-calls

About LFM3-1.7B

LFM3-1.7B is the text model in LFM3, the third generation of Liquid Foundation Models from Liquid AI. Liquid announced the family on 27 October 2026 and published this checkpoint, post-trained from LFM3-1.7B-Base, as one of six open-weight models of about 2B parameters. The base was pre-trained on 22T tokens, and the last 3.5T were curated mid-training data rich in agentic traces and tool calls. Liquid's deprecations page now lists it as the recommended replacement for LFM2.5-2.6B, and the announcement says it beats that model while using 35% fewer parameters. The earlier LFM line is represented on Hokai by the vision model LFM2.5-VL-3B. More from this maker sits in the Liquid AI model directory, and similar releases are grouped under open-weight models.

Every score here is vendor-reported. It comes from the table on Liquid's model card, where all models were served in BF16 with vLLM at their full context length, and no independent evaluator lists LFM3-1.7B yet. On agents and tool use Liquid reports 71.39 on BFCL V4, 80.36 on ToolSandbox, 38.42 on WebArena-Verified and 45.60 on LiveClaw, ahead of the 9.7B Qwen3.5-9B on each (62.83, 76.41, 28.94 and 35.77). It also shows 73.03 on IFBench, 73.46 on LiveCodeBench v6 and a 49.37 average across all rows against 47.24 for Qwen3.5-9B. The larger model keeps the lead on MCP-Atlas (42.67 against 32.57), GPQA Diamond (80.40 against 63.37), Humanity's Last Exam without tools (16.41 against 7.17), SWE-bench Verified with the OpenCode agent setup (51.20 against 43.40) and Terminal-Bench 2.1 (17.98 against 12.92).

Liquid says the model is not recommended for agentic coding or complex multi-step reasoning, and that knowledge-heavy tasks need retrieval or search; with search its Humanity's Last Exam score nearly doubles, from 7.17 to 13.90. The AA-Omniscience rows show the same pattern from another side: 7.07 accuracy, but a 92.93 non-hallucination score, against 13.17 and 52.59 for Qwen3.5-9B. Those figures reconcile with a model that declines far more often than it answers wrongly, which suits an agent that can fall back to a search tool. Long context is mixed: 65.56 on RULER 128K trails the 84.63 of the 4.7B Qwen3.5-4B, while the two tie on AA-LCR (52.83 and 52.67).

Liquid post-trained the model inside agent harnesses such as Hermes Agent and OpenClaw. Its guide to running those locally is written around LFM2.5-2.6B, so the commands there need the LFM3 model id and a recent llama.cpp. In a testimonial printed in the announcement, MacPaw's LLM efficiency lead says LFM3-1.7B beat Qwen3.5-2B on 65 everyday agent tasks in real macOS environments, with 64% task success against 48% and 58% fewer tokens per task. That is one customer's test, published without a method. Hokai's guide Hermes Agent vs OpenClaw compares the two frameworks.

GPU serving runs through Transformers, vLLM and SGLang. On CPUs and phones, llama.cpp is the documented route, with builds on Hugging Face in BF16, FP8, NVFP4, GGUF, MLX and ONNX. Ollama and LM Studio also load GGUF files, subject to the runtime-version quirk below, and the guides How to Run AI Locally in 2026 and Ollama vs LM Studio cover choosing between them. Hosted access comes from Liquid Console and a Playground. Liquid names OpenRouter and Vercel as launch partners, but neither public model catalog listed any LFM3 model on 4 November 2026, so the Liquid API is the one verified hosted route.

Pricing

USD per million tokens as printed on Liquid's pricing page on 4 November 2026: $0.08 input, $0.02 cached input, $0.20 output. That page prices LFM3-VL-2B, LFM3-Audio-2B and LFM3-Omni-2B separately, so match the model name before comparing rates. Self-hosting adds no license fee, and a company past the license's revenue limit gets its quote from Liquid sales. For an alternative with a published API rate, see [Ministral 3 3B](/hub/models/ministral-3-3b); the family overview is at [LFM3](/hub/models/lfm3).

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Tri-hybrid layer mix: The 24 layers combine 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers, a layout Liquid searched for on 14 device and backend combinations.
• Cheap long context: Liquid measures 1.6 GB of KV cache for a full 128K-token context, against 4.3 GB for Qwen's 4.7B model and 5.6 GB for MiniCPM5-2B.
• Deferred tools and tool search: Tools marked defer_loading stay out of the prompt until a search tool returns their names, so one agent can choose from a large tool catalog without paying for every schema up front.
• Small quantized files: The quantization-aware-distilled Q4_0 GGUF is 1.05 GB and the Q4_K_M build is 1.10 GB, against 3.46 GB for the BF16 safetensors.
• Single-GPU fine-tuning recipes: Liquid ships TRL notebooks for SFT, DPO, GRPO and continued pre-training, including a GRPO run over a 108-tool phone app and a Turkish language example.

Pros

• Beats a model more than five times its size on function calling and browser-style agent tasks in the vendor's own comparison.
• Abstains more than it guesses on the AA-Omniscience public subset, which is a useful trait when an agent can call a search tool.
• Small enough to fine-tune on one GPU, so a narrow in-house specialist is realistic for a small team.

Cons

• Weak where size matters: the vendor itself advises against agentic coding and complex multi-step reasoning.
• The LFM Open License v1.0 is not Apache 2.0: commercial use is licensed only below $10 million in annual revenue, counting parents and affiliates, unlike Ministral 3 3B.
• No safety evaluation or system card accompanies the release, and no independent benchmark site lists the model yet.

Benchmarks

• IFBench: 73.0% vendor-reported · 04 Nov 2026 — How precisely the model follows detailed instructions, % passing.
• MCP Atlas: 32.6% vendor-reported · 04 Nov 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
• GPQA Diamond: 63.4% vendor-reported · 04 Nov 2026 — PhD-level science questions that are hard to search for, % correct.
• LiveCodeBench v6: 73.5% vendor-reported · 04 Nov 2026 — Recently published coding problems, % solved.
• SWE-bench Verified: 43.4% vendor-reported · 04 Nov 2026 — Real GitHub issues fixed end to end, % solved.
• Terminal-Bench 2.1: 12.9% vendor-reported · 04 Nov 2026 — Multi-step tasks completed in a real command line, % solved.
• Humanity's Last Exam: 7.2% vendor-reported · 04 Nov 2026 — Expert-written questions across many fields, % correct.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit LFM3-1.7B Official Page