Liquid AI Releases LFM2.5-VL-3B, a 3.1B-Parameter Vision-Language Model for On-Device Inference
Liquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model built for on-device deployment. The model claims leading performance in its size class on grounding, screen understanding, and tool use, while running at 228 tokens/s on an Apple M5 Max.
How LFM2.5-VL-3B ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| MMMU | 48.4% | 12 of 13 | Hugging Face model card ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Liquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model designed to run on local hardware rather than in the cloud. The model is available now on Hugging Face and ships with support for llama.cpp, MLX, vLLM, SGLang, and ONNX.
What's new
LFM2.5-VL-3B builds on Liquid AI's earlier LFM2-VL-3B with four claimed improvements: screen and UI understanding across device types, improved object grounding via natural-language queries, better multi-image reasoning, and significantly stronger function calling in both text-only and vision-text settings. The model answers directly rather than using chain-of-thought reasoning, which Liquid AI says keeps latency low for real-time and on-device use.
Architecture and training
The model pairs a SigLIP2 400M NaFlex vision encoder with the same pre-trained backbone used in Liquid AI's LFM2.5-2.6B text model. According to Liquid AI, it was pre-trained on approximately 34 trillion tokens, including 4x more vision data than prior releases, drawn from curated and synthetic image-caption, OCR, grounding, and instruction-following datasets. The tokenizer vocabulary was expanded to 128K to support non-Latin scripts, done by extending the existing tokenizer rather than retraining from scratch.
Post-training used two stages: supervised fine-tuning with knowledge distillation from a larger teacher model plus a technique Liquid AI calls "Antidoom training," followed by multi-reward reinforcement learning.
Benchmark results (as reported by Liquid AI)
Across vision benchmarks, LFM2.5-VL-3B posted an average score of 69.4, ahead of Gemma-4-E2B-it (57.2, 5.1B params), InternVL 3.5 2B (64.6), and Qwen3.5-2B (63.7), though behind the larger Qwen3.5-4B (70.1, 4.7B params). Notable individual scores include MathVista-mini (68.5), DocVQA-val (91.1), and RefCOCO-avg grounding (87.9, versus 57.1 for its predecessor LFM2-VL-3B).
On text-only benchmarks, tool-use performance improved sharply: BFCL V4 scored 32.5 versus 20.5 for the prior version, and ToolSandbox scored 59.5 versus 26.4. Liquid AI says this puts the model roughly on par with Gemma-4-E2B and Qwen3.5-2B on tool use. Instruction-following scores (IFEval: 82.3, Multi-IF: 59.4) also rose over the predecessor model.
Inference speed
Liquid AI reports the model decodes at 228 tokens/second on an Apple M5 Max and 116 tokens/second on a Ryzen AI Max+ 395, fitting in roughly 3 GB of memory. On a Galaxy S26 Ultra, it reportedly reaches 20 tokens/second running fully on-device. On GPU, the company claims output throughput of about 11,000 tokens/second at high concurrency on a single H100 — roughly double that of larger 4B-class models — translating to nearly 1 billion output tokens per day per GPU, according to Liquid AI's testing.
Access
The model is available on Hugging Face under the identifier LiquidAI/LFM2.5-VL-3B, compatible with transformers>=5.10.1. Liquid AI has also published a browser-based WebGPU demo supporting multi-image chat, grounding, OCR, and tool calling. Pricing is not applicable; the model is distributed as open weights for local deployment. Training data cutoff was not disclosed.
What this means
LFM2.5-VL-3B targets a specific niche: vision-language capability that fits on phones, laptops, and edge devices rather than requiring cloud inference. The reported gains in grounding and tool calling matter for agentic use cases — apps that need to identify UI elements and act on them, such as screen automation or GUI agents. The claimed throughput advantage on GPU (11K tokens/second at high concurrency) also positions the model for high-volume server deployment, not just edge use, if Liquid AI's benchmarks hold up under independent testing. As with any vendor-reported benchmark, the comparisons against Gemma, InternVL, and Qwen model families should be treated as claims until third parties replicate them.
Related Articles
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Paradigm Intelligence launches Phai-A1 language model and Phai-M1 video model, trained on Huawei Ascend
Paradigm Intelligence (范式智能, formerly 第四范式) has unveiled two in-house models: Phai-A1, a language model that runs on a single RTX 3090, and Phai-M1, a video model with synchronized audio. The company says both were trained on Huawei Ascend hardware, with open weights coming soon.
Cloudflare releases open-weight Clairic 27B and new Clairic-flash 9B for data extraction
Cloudflare announced on November 18, 2026 an updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash. Both have open weights on Hugging Face and are available as hosted Workers AI endpoints.
RadixArk releases DaisyRose-27B, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on 1,200 curated agent tasks. According to RadixArk, it raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than the base model on all seven reported benchmarks. Weights ship under Apache 2.0 in BF16, FP8 and NVFP4.
Comments
Loading...