Liquid AI released the LFM3 family at the end of October, with its most unassuming member LFM3-1.7B Instead, the most compelling option: 1.73B parameters, 256K context, ~1GB memory footprint after 4-bit quantization, launched on OpenRouter on November 7, with input cost of just $0.08 per million tokens.
Pitted against Qwen3.5-9B—five times its size—its Agent benchmark scores are LiveClaw 45.60 vs. 35.77 and AutomationBench 41.29 vs. 15.10; for instruction following, IFEval 96.66 ranks it first overall. How can a model small enough to fit on a smartphone outperform a 9B model in Agent scenarios?
What exactly is it: Tri-hybrid architecture + Agent-specific post-training
LFM3 is Liquid AI’s next-generation family, comprising six product lines: base, text, vision, audio, omni, and decision. LFM3-1.7B is the pure-text variant, built on a new tri-hybrid architecture.
- Parameter countParameters: 1.73B, vocabulary size: 128,000
- ArchitectureArchitecture: 24 layers = 13 layers of gated short convolutions + 5 layers of sliding-window attention + 6 layers of global attention. Convolutions handle local information efficiently; only 6 layers of global attention manage long-range dependencies—this is how it fits 256K context into just 1GB of memory.
- contextContext length: 256K tokens (input + output); maximum single-output length on OpenRouter is 8,192 tokens
- LanguageLanguages supported: Officially lists 16, including Chinese, Japanese, Korean, Arabic, Vietnamese, and Thai
- Post-training focusExplicitly reinforced for Agent harnessing (officially named: Hermes Agent, Pi, OpenClaw), tool calling, and structured output generation
- LicenseLFM Open License v1.0; weights openly available for download on Hugging Face, with over 70,000 downloads to date
Training orientation in one sentence: Of the 22 trillion tokens trained on, the final 3.5 trillion were replaced with high-quality Agent trajectories and tool-calling data, then distilled back into a single model via online policy distillation with six expert teachers (MOPD). It is not “capable enough to be used as an Agent”—it was trained *as* an Agent from late-stage training onward.
Benchmark performance: An outlier among small models—but with pronounced specialization
Data sourced from the official model card, benchmarked using vLLM under full BF16 context. For comparison, Qwen3.5-4B and Qwen3.5-9B are both larger than it.
| Benchmark | LFM3-1.7B | LFM2.5-2.6B | Qwen3.5-4B | Qwen3.5-9B |
|---|---|---|---|---|
| LiveClaw (Agent comprehensive) | 45.60 | 14.70 | 19.94 | 35.77 |
| AutomationBench | 41.29 | 0.94 | 8.67 | 15.10 |
| WebArena-Verified | 38.42 | 3.02 | 21.92 | 28.94 |
| BFCL V4 (function calling) | 71.39 | 55.89 | 55.12 | 62.83 |
| IFEval (instruction following) | 96.66 | 95.39 | 92.52 | 92.08 |
| GPQA Diamond (knowledge reasoning) | 63.37 | 57.78 | 75.76 | 80.40 |
| SWE-bench Verified (programming) | 43.40 | 18.00 | 41.20 | 51.20 |
| RULER 128K (long-context retrieval) | 65.56 | 1.15 | 84.63 | 87.11 |
The conclusion is clear:It leads comprehensively on tasks that follow a defined workflow, but lags noticeably on tasks requiring pure mental effort.Average across 24 benchmarks: 49.37—higher than Qwen3.5-9B’s 47.24, driven by gains in Agent and instruction-following tasks.
More notably, “no hallucination”: accuracy in AA-Omniscience is only 7.07%, yetNon-hallucination rate is 92.93%, the highest among all models.Small models knowing their own limits is more valuable in tool-calling pipelines than memorizing obscure facts.
User experience and real limitations
Deployment has almost no barrier: vLLM, llama.cpp, Ollama, and MLX—GGUF with Q4_0 quantization suffices. Official benchmarks show it pre-fills at least 2.5× faster and decodes 1.8× faster than alternatives under 32K context on Ryzen AI Max+ 395, M5 Max, and iPhone 17 Pro, while using the least memory.
But hard limitations are explicit: the official model card itself states it is not recommended for agentic coding or complex multi-step reasoning. Data confirms this—Terminal-Bench 2.1 score is only 12.92; MCP-Atlas is 32.57 (Qwen3.5-9B scores 42.67).
- Thin knowledge baseHLE: 7.17% without tools, 13.90% with tools—retrieval remains essential for domain-specific details.
- Long-context: “can ingest, cannot chew”RULER 128K score is only 65.56; ingesting 200,000 characters does not guarantee reliable extraction—RAG integration is advised for retrieval tasks.
- Chinese: “supported”, not “proficient”Short instructions and format conversion are stable; Chinese long-text writing fluency lags behind similarly priced domestic models
- Not suitable as a primary programming modelScript writing and configuration changes work fine, but don’t expect cross-file refactoring
Pricing and access methods
Three options: OpenRouter’s fp8 endpoint—pay-per-use at $0.08 per million input tokens and $0.20 per million output tokens, cached reads at $0.02, plus a free variant for testing; pull weights from Hugging Face for local deployment—4-bit quantization requires ~1 GB; or perform domain-specific fine-tuning on a single GPU, which the official team emphasizes as the core value proposition for small models
LFM3-1.7B vs. other models in horizontal comparison
| Model | Parameter count | Core Positioning | This Site’s Rating |
|---|---|---|---|
| LFM3-1.7B | 1.73B / 256K | On-device Agent and tool calling | 8.4 |
| Liquid LFM 2.5 | 24B (A2B sparse) | Previous-generation Agent inference model | 8.0 |
| Qwen3.8 2.4T A95B | 2.4T MoE | Domestic open-source flagship | 8.8 |
| Wenxin 5.5 Flash Base | 280B MoE | High-value open-source text model | 8.3 |
| Nemotron 3.5 Super VL | 120B | Enterprise private deployment | 8.5 |
Position is clear: absolute capability falls short of the top tier, but it has no rival in “Agent capability per GB of memory or per dollar”—it doesn’t replace flagship models, only offloads frequent lightweight tasks
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 8.3 / 10 | Full 256K context, tool calling, and structured output support—but weak MCP ecosystem integration |
| 易用性 | 9.2 / 10 | Runs on just 1 GB RAM; fully supports GGUF/MLX/Ollama/vLLM—even on smartphones |
| Cost-effectiveness | 9.3 / 10 | $0.08 per million input tokens; zero cost for local deployment; plus a free variant |
| 中文支持 | 7.8 / 10 | Official Chinese support; short-instruction stability; average long-text writing fluency |
| 输出质量 | 7.5 / 10 | Top-tier instruction following, no hallucination—but shallow knowledge base and weak complex reasoning |
Overall rating: 8.4/10
Frequently Asked Questions (FAQ)
Is LFM3-1.7B free?
Weights are freely downloadable under the LFM Open License v1.0; running locally only incurs electricity costs. API usage is pay-as-you-go: $0.08 per million input tokens and $0.20 per million output tokens on OpenRouter, plus a rate-limited free variant.
What hardware configuration is required for local deployment?
4-bit quantization requires ~1 GB of memory—plenty for a 16 GB laptop; officially tested to run even on iPhone 17 Pro. For full 256K context length, ≥8 GB GPU memory is recommended; KV cache also consumes memory.
Is it suitable for coding?
It handles single-file scripts, config edits, and structured output generation well—but not cross-file refactoring. Officially, it is explicitly not recommended for Agentic coding. For serious coding, see Kimi K2.7 Code or GLM 5.3.
Which should I choose: LFM3-1.7B or Qwen3.5-9B?
Choose by task type: LFM3-1.7B excels at tool calling, web interaction, and workflow automation—and costs over ten times less; Qwen3.5-9B is better for knowledge Q&A, complex reasoning, and long Chinese text. The most cost-efficient approach is to integrate both with routing.
Does it support Chinese? How good is the performance?
Yes—it supports Chinese and 15 other languages per the official list. Short instructions and format conversion are stable; however, long-form Chinese content generation is noticeably weaker than similarly priced domestic models—switch models for such tasks.
Summary: A small model’s value lies beyond benchmark scores
Its significance is not clickbait claims like “1.7B beats 9B,” but rather slashing the cost of high-frequency, low-value, no-brain tasks inside Agents—parameter parsing, routing decisions, format conversion, status monitoring—to near-negligible levels. Flagship models think; this one executes. This is today’s most pragmatic division of labor.
Want to Discover More Useful AI Tools? Explore Our AI Model Library and Tool Comparison Engineor continue reading:Liquid LFM 2.5 Review · Tencent Hunyuan Hy4 Review · Document your initial hypothesis and metric selection with a timestamp. Experiments show this does not fully prevent influence, but at least lets you see afterward what changed in your thinking.