Model Card | Liquid AI LFM3-1.7B

🌏 中文版

Model Information

Capability Highlights

• Scores 38.42 on WebArena-Verified, beating the 5x-larger Qwen3.5-9B (28.94); also beats it on the LiveClaw agent benchmark, 45.60 vs 35.77
• Uses 35% fewer parameters than its predecessor LFM2.5-2.6B (1.73B vs 2.6B) while beating it on every benchmark Liquid AI published side by side
• On three test devices (AMD Ryzen AI Max+ 395, Apple M5 Max, iPhone 17 Pro) at 32K context, it prefills at least 2.5x faster and decodes at least 1.8x faster than every compared model, while using the least memory; at 4-bit precision it needs only about 1GB
• The tri-hybrid architecture cuts KV cache to roughly 27% of a standard 24-layer transformer, which is how the same memory budget stretches to a 256K context; total pre-training was 22T tokens

Benchmark Results

⚠️ All figures above are Liquid AI's own benchmarks (vLLM, BF16, run at full context length); no independent third-party reproduction exists as of release. Multiple outlets (VentureBeat, TechTimes) also flag these as vendor-reported numbers.

Comparison with Predecessor and Competitors

Against its predecessor LFM2.5-2.6B, LFM3-1.7B isn't just smaller — the architecture changed entirely: the tri-hybrid design (13 short-convolution, 5 sliding-window attention, 6 global attention layers) replaces what came before, and pre-training data grew to 22T tokens. The result is a generational jump rather than a tune-up: the predecessor could barely answer WebArena-Verified tasks (3.02), while LFM3-1.7B scores 38.42.

Against Qwen3.5-9B — five times the parameter count — LFM3-1.7B wins across every "hands-on" agent task (WebArena-Verified, LiveClaw, BFCL V4 tool calling), and edges ahead on the unweighted average too (49.37 vs 47.24). But it still loses on long-context understanding (AA-LCR, 52.83 vs 62.00) and knowledge-heavy benchmarks (GPQA Diamond, 63.37 vs 80.40; Humanity's Last Exam trails across the board). That profile shows LFM3-1.7B is deliberately optimized for "doing" rather than "knowing."

Pricing also shifted noticeably: the Liquid Console API's $0.08/1M input and $0.20/1M output is far cheaper than the predecessor LFM2.5-2.6B's $0.35/$0.50, as tracked by third-party aggregators (VentureBeat cites the official console rate card, which requires login to view; the LFM2.5-2.6B figures come from pricepertoken.com). A smaller model at a lower price, with free on-device deployment on top of open weights, is the clearest commercial signal in this release.

What This Means for Agent Development

The tri-hybrid architecture's ~27% KV cache footprint is why Liquid AI could push the text model's context window all the way to 256K on the same memory budget. That matters most for "always-on" local agents — ones running continuously on a phone or laptop and accumulating conversation history — since they don't need to clear context frequently or lean on a cloud API for memory.

• If you're building an always-on local or enterprise-controlled agent (data stays on-device): LFM3-1.7B needs roughly 1GB of memory at 4-bit precision, ships with a 256K context, and officially supports agent harnesses like Hermes Agent, Pi, and OpenClaw. The release names Mercedes-Benz (exploring it for the MBUX in-car assistant) and Shopify as pre-release testing partners
• If you're running high-concurrency cloud tool-calling services: a single NVIDIA B300 can produce 3.6 billion tokens/day at 1,024 concurrent requests, and it beats Qwen3.5-9B on several agentic benchmarks with a fifth of the parameters — worth evaluating for cost-sensitive workloads
• Not a good fit: Hugging Face's own model card explicitly says it isn't recommended for agentic coding or complex multi-step reasoning (SWE-bench Verified is only 43.40, trailing Qwen3.5-9B's 51.20), or for knowledge-heavy tasks without retrieval

Today's Takeaway

The interesting part here isn't the familiar "small model beats big model" story — it's how clearly Liquid AI spells out the trade-off. LFM3-1.7B beats a model five times its size on tool-use and web-navigation tasks, at the cost of falling behind on long-context understanding and knowledge-heavy questions, and the official model card itself discourages using it for agentic coding. The same release also ships two "decision models" (the d1 series) that return structured answers instead of conversational text — a reminder that not every agent task needs a model that can chat; some just need a reliable classifier.

References

• Liquid AI Blog: Introducing LFM3 — Every Modality for Specialized Intelligence (official release post)
• Official HuggingFace model card: LiquidAI/LFM3-1.7B
• VentureBeat: Liquid AI's new LFM3 models put browser agents, voice commands, and image recognition onto enterprise-controlled devices (includes Liquid Console API pricing)
• TechTimes: Liquid AI Releases LFM3 — Six Open-Weight Models Fuse Text, Vision, Audio, One 2B Backbone (tri-hybrid architecture and KV cache details)
• startupfortune.com: Liquid AI releases LFM3 models small enough to run on a phone (LFM3-VL-2B / LFM3-Audio-2B details)
• pricepertoken.com: Liquid AI API Pricing (LFM2.5-2.6B pricing comparison)
• Our Liquid AI / LFM family primer
• Our model card: Liquid AI LFM2.5-24B-A2B (previous entry)