Liquid AI releases LFM3 models small enough to run on a phone

Liquid AI released LFM3 on October 27, a family of six open-weight models built on a shared 2-billion-parameter backbone, including LFM3-1.7B, which runs with a 256,000-token context window and fits in roughly 1GB of memory at four-bit precision. According to MarkTechPost's review of the release, LFM3-1.7B averages 49.37 across 29 benchmarks, ahead of Alibaba's Qwen3.5-9B at 47.24, and scores 41.29 on AutomationBench versus 15.10 for Qwen3.5-9B. The release lands as Nvidia Blackwell rental prices have jumped 48% in two months and TSMC's CoWoS packaging is booked through at least mid-2027, positioning on-device models as an alternative to scarce GPU capacity.

Liquid AI shipped six open-weight models on October 27 that fit in about a gigabyte of memory, right as Blackwell GPU rentals are up 48% in two months and training queues stretch into 2027. Liquid AI announced LFM3 on October 27, a new family of six open-weight models built around a shared 2-billion-parameter backbone, according to the company's own announcement and reporting from MarkTechPost. The lineup spans text, vision, audio, and decision-making, and the whole point is that none of it needs a datacenter. LFM3-1.7B, the text model, runs with a 256,000-token context window and fits in roughly 1GB of memory at four-bit precision, small enough for a phone or a laptop with no GPU in sight. That timing is not an accident. Nvidia Blackwell rental prices have jumped 48% in two months, and enterprises that didn't lock in orders by early 2026 are now staring at Q1 2027 delivery windows, according to industry tracking cited by Clarifai and GPUNEX. TSMC's CoWoS packaging is fully booked through at least mid-2027, and SK Hynix's HBM memory is reserved 12 to 15 months out across the only three suppliers who make it. If you're a founder building an AI product right now, you are either waiting in that line or looking for a way around it. LFM3 is Liquid AI's bet that there is a way around it. The architecture behind it is what Liquid AI calls tri-hybrid: 24 layers split between 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers, found through what the company describes as a hardware-in-the-loop search across 14 device and backend combinations, including a Raspberry Pi 5 and a Qualcomm Hexagon NPU. That's not a theoretical efficiency claim. It's an architecture tuned against the actual chips a phone or a laptop ships with, rather than against a GPU cluster that most startups can't afford to rent anymore. Small is easy. Competitive is the harder claim, and Liquid AI backs it with numbers. LFM3-1.7B averages 49.37 across 29 benchmarks, ahead of Alibaba's Qwen3.5-9B at 47.24, a model more than five times its size, according to MarkTechPost's review of the release. On IFEval, a test of how well a model follows instructions, LFM3-1.7B scores 96.66. On AutomationBench, a tool-use and agentic benchmark, it hits 41.29 against 15.10 for Qwen3.5-9B. The vision model, LFM3-VL-2B, is Liquid's first to accept video input and averages 71.4 across 62 benchmarks, up from 57.6 for its own predecessor, LFM2.5-VL-3B, despite having fewer parameters. Builders Running AI in Production Are Quietly Betting Smaller Beats Bigger https://startupfortune.com/builders-running-ai-in-production-are-quietly-betting-smaller-beats-bigger/ Microsoft's Phi-4 beats GPT-4o on the MATH and GPQA benchmarks despite having just 14 billion parameters. Frontier API pricing runs $15 to $75 per million tokens, a gap that can make small, fine-tuned models far cheaper to run at scale. - deploying smaller language models in production https://startupfortune.com/builders-running-ai-in-production-are-quietly-betting-smaller-beats-bigger/ - running local AI models versus frontier models https://startupfortune.com/builders-running-ai-in-production-are-quietly-betting-smaller-beats-bigger/ The audio model, LFM3-Audio-2B, supports 18 languages and posts a speech-quality score of 3.83 on UTMOSv2 against 3.22 for the prior LFM2.5-Audio-1.5B, with a causal encoder that gets time-to-first-token down to as low as 50 milliseconds. There's also LFM3-Omni-2B, which handles multiple modalities at once, and two decision models aimed at agentic tasks rather than conversation. Six models, one shared backbone, each one specialized rather than general-purpose. That specialization is the real pitch to builders, not just the file size. Liquid AI says the models can be adapted on a single GPU, which matters enormously for a startup that has one H100 and a deadline, not a reservation in Nvidia's 2027 delivery queue. Weights are downloadable from Hugging Face. Arm publicly backed the launch, saying it has worked with Liquid AI to get LFM3 running efficiently on Arm chips, the architecture inside most phones and a growing share of laptops. None of this makes the cloud compute crunch disappear. Training genuinely large frontier models still needs the GPUs that are sold out, and nothing about a 2B-parameter phone model changes that math for the labs racing toward bigger systems. What it does is give the rest of the market, the startup that wants an AI feature without a six-figure cloud bill, a real alternative that doesn't involve waiting in a queue behind Microsoft and OpenAI. Liquid AI still competes in a crowded field here: Google's Gemma, Mistral's small models, and Alibaba's Qwen line are all chasing the same efficient, on-device niche, and Qwen3.5-9B is the exact model LFM3 benchmarks itself against. Liquid AI raised $250 million in a Series A in December 2024 at a $2.35 billion valuation, led by AMD Ventures and Samsung Catalyst Fund, money that reads differently now than it did then. Back then, on-device efficiency was a nice architectural story. In October 2026, with HBM booked more than a year out and GPU rental prices climbing by the month, it's starting to look like the plan. Also read: SpAItial launches Echo-3 to turn iPhone scans into 3D digital twins https://startupfortune.com/spaitial-launches-echo-3-to-turn-iphone-scans-into-3d-digital-twins/ • ASMPT sales jump 49% on AI chip packaging as core equipment unit misses https://startupfortune.com/asmpt-sales-jump-49-on-ai-chip-packaging-as-core-equipment-unit-misses/ • Wall Street keeps betting on IREN's pivot from bitcoin miner to AI cloud https://startupfortune.com/wall-street-keeps-betting-on-irens-pivot-from-bitcoin-miner-to-ai-cloud/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.