LFM2.5-VL-3B

--- title: “LFM2.5-VL-3B” description: “LFM2.5-VL-3B (Liquid AI, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-lfm2.5-vl-3b.png imageAlt: “A retro robot representing LFM2.5-VL-3B” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Liquid AI blog — LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge (Aug 12, 2026)” url: “https://www.liquid.ai/blog/lfm2-5-vl-3b” - name: “Hugging Face model card — LiquidAI/LFM2.5-VL-3B” url: “https://huggingface.co/LiquidAI/LFM2.5-VL-3B” - name: “Liquid AI Docs — LFM2.5-VL-3B model page” url: “https://docs.liquid.ai/lfm/models/lfm25-vl-3b” - name: “unite.ai — Liquid AI Ships LFM2.5-VL-3B for Faster Vision-Language AI on the Edge (Aug 12, 2026)” url: “https://www.unite.ai/liquid-ai-ships-lfm2-5-vl-3b-for-faster-vision-language-ai-on-the-edge/” - name: “VentureBeat — No cloud, no GPUs, no problem: Liquid AI’s new model LFM2.5-2.6B (Aug 6, 2026)” url: “https://venturebeat.com/technology/no-cloud-no-gpus-no-problem-liquid-ais-new-model-lfm2-5-2-6b-brings-powerful-ai-agents-to-devices-as-small-as-a-raspberry-pi” - name: “Unsloth — LFM2.5 run + fine-tune guide” url: “https://unsloth.ai/docs/models/tutorials/lfm2.5” - name: “alphasignal.ai — Liquid AI’s LFM2.5-VL-3B beats models twice its size, running on your phone (Aug 12, 2026)” url: “https://alphasignal.ai/news/liquid-ai-s-lfm2-5-vl-3b-beats-models-twice-its-size-running-on-your-phone” facts: - label: “Vendor” value: “Liquid AI” - label: “Released” value: “2026” - label: “License” value: “Open-weight (Liquid)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - lfm2.5-vl-3b - ai-models - model-release draft: false --- --- title: “LFM2.5-VL-3B” description: “LFM2.5-VL-3B (Liquid AI, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-lfm2.5-vl-3b.png imageAlt: “A retro robot representing LFM2.5-VL-3B” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Liquid AI blog — LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge (Aug 12, 2026)” url: “https://www.liquid.ai/blog/lfm2-5-vl-3b” - name: “Hugging Face model card — LiquidAI/LFM2.5-VL-3B” url: “https://huggingface.co/LiquidAI/LFM2.5-VL-3B” - name: “Liquid AI Docs — LFM2.5-VL-3B model page” url: “https://docs.liquid.ai/lfm/models/lfm25-vl-3b” - name: “unite.ai — Liquid AI Ships LFM2.5-VL-3B for Faster Vision-Language AI on the Edge (Aug 12, 2026)” url: “https://www.unite.ai/liquid-ai-ships-lfm2-5-vl-3b-for-faster-vision-language-ai-on-the-edge/” - name: “VentureBeat — No cloud, no GPUs, no problem: Liquid AI’s new model LFM2.5-2.6B (Aug 6, 2026)” url: “https://venturebeat.com/technology/no-cloud-no-gpus-no-problem-liquid-ais-new-model-lfm2-5-2-6b-brings-powerful-ai-agents-to-devices-as-small-as-a-raspberry-pi” - name: “Unsloth — LFM2.5 run + fine-tune guide” url: “https://unsloth.ai/docs/models/tutorials/lfm2.5” - name: “alphasignal.ai — Liquid AI’s LFM2.5-VL-3B beats models twice its size, running on your phone (Aug 12, 2026)” url: “https://alphasignal.ai/news/liquid-ai-s-lfm2-5-vl-3b-beats-models-twice-its-size-running-on-your-phone” facts: - label: “Vendor” value: “Liquid AI” - label: “Released” value: “2026” - label: “License” value: “Open-weight (Liquid)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - lfm2.5-vl-3b - ai-models - model-release draft: false --- --- type: guide category: ai model: lfm2.5-vl-3b vendor: Liquid AI release_date: 2026-08-12 status: current license: LFM Open License v1.0 (open-weight, free for orgs under $10M ARR) --- # LFM2.5-VL-3B — Liquid AI’s 3B on-device vision-language model that runs on a phone ## The release Liquid AI shipped LFM2.5-VL-3B on August 12, 2026, the vision twin of the LFM2.5-2.6B text model the Cambridge, MA-based MIT spinoff released eight days earlier. It is the largest of Liquid’s three current VL sizes — 450M, 1.6B, 3B. The official blog frames it bluntly: LFM2.5-VL-3B decodes 228 tokens/s on an Apple M5 Max and 116 tokens/s on an AMD Ryzen AI Max+ 395, while staying in about 3 GB of memory. Same release confirmed the open-weight posture: Liquid AI’s most capable vision-language model…delivers competitive vision performance against models twice its size — Liquid AI blog. LFM2.5-VL-3B pairs the LFM2.5-2.6B text backbone with the SigLIP2 NaFlex 400M image encoder and a lightweight projector — a dense 3.1B-parameter model, not mixture-of-experts. Pretrained on ~34T tokens, with vision pretraining scaled 4× versus the predecessor LFM2-VL-3B. Same 32K context ceiling as the family, same 128K tokenizer. License: LFM Open License v1.0 — free for individuals, researchers, and companies with < $10M annual revenue, commercial license above. Day-one surface: raw weights + GGUF, ONNX, and MLX quantizations on the same HF repo; llama.cpp, MLX, vLLM, SGLang, and ONNX runtimes all supported; Unsloth + TRL fine-tuning recipe already published. alphasignal.ai’s headline within hours: Liquid AI releases LFM2.5-VL-3B, a 3.1B vision-language model that beats models up to 2.6× its size on screen understanding and grounding. (alphasignal.ai) ## Why it matters for operators This is the first time a 3B-class VLM is plausibly competitive with 5–8B open-weight models on a meaningful fraction of the screen-and-document work, while actually running on a phone. unite.ai’s framing is the right one: Liquid AI released LFM2.5-VL-3B on August 12, 2026, a 3.1-billion-parameter open-weight vision-language model built to run on phones, laptops, and single GPUs rather than in a data center. (unite.ai, Aug 12, 2026) The phone-decode number — 20 tok/s on a Galaxy S26 Ultra — is the one ABS operators should care about: fast enough for a real “snap product, get a caption” loop without sending the image to a third party. ## Specs that matter | Spec | Value | |---|---| | Architecture | Dense LFM2.5 hybrid backbone + SigLIP2 NaFlex image encoder + projector | | Total parameters | 3.1B (≈2.6B language + ~400M vision) | | Context length | 32,768 tokens | | Tokenizer | 128K vocabulary (doubled for non-Latin scripts) | | Modes | Non-reasoning only (by design, for low latency) | | Inputs | Text + image(s) — multi-image supported | | Outputs | Text | | Runtime memory | ~3 GB claimed for on-device inference | | On-device speed | 228 tok/s (M5 Max) · 116 tok/s (Ryzen AI Max+) · 20 tok/s (Galaxy S26 Ultra) · ~11K tok/s aggregate on H100 (vLLM 0.26) | | License | LFM Open License v1.0 (free <$10M ARR; commercial license above) — LFM2.5-2.6B is distributed under the LFM Open License v1.0, which permits use, modification, and redistribution — including commercial use — for organizations with less than $10 million in annual revenue. (VentureBeat) | | Formats | Raw weights + GGUF + ONNX + MLX quantizations, day one | | Runtimes | llama.cpp, MLX, vLLM, SGLang, ONNX (day-one support) | Docs page confirms the architecture in plain prose: Liquid AI’s most capable vision-language model, a SigLIP2 NaFlex image encoder, Context Length 32K tokens, Architecture LFM2.5-VL (Dense). (Liquid AI Docs — LFM2.5-VL-3B) Not in scope: native video understanding — multi-frame 5×256×256 clips are accepted in the GPU benchmark, but the model is positioned as a still-image VLM with multi-image support, not as a video model. ## Benchmarks All numbers below are vendor-reported — Liquid AI ran them on vLLM 0.26 in non-reasoning mode. Treat them as the official claim, not an independent audit: All benchmark and speed figures in this release are vendor-reported. Average across Liquid’s 28-benchmark suite: LFM2.5-VL-3B 69.4 · LFM2-VL-3B predecessor 57.2 · Gemma-4-E2B-it 52.0 · Gemma-4-E4B-it 59.7 · InternVL 3.5 2B 64.6 · InternVL 3.5 4B 69.4 · Qwen 3.5 2B 63.7 · Qwen 3.5 4B 70.1. → LFM2.5-VL-3B ties InternVL 3.5 4B on average (4.7B vs 3.1B) and trails Qwen 3.5 4B by 0.7 points at roughly 65% of the parameters. Selected vision scores (LFM2.5-VL-3B): - MMBench EN 81.0 · RealWorldQA 73.1 · POPE 88.7 - DocVQA 91.1 · ChartQA 81.3 · TextVQA 84.3 · OCRBench v1 84.2 · OCRBench v2 EN 47.5 · InfographicVQA 70.2 - ScreenSpot-v2 average 80.7 (Desktop 78.7 / Mobile 81.2 / Web 82.2) — the standout gain from predecessor’s 6 / 7.6 / 2.5 - Object grounding: RefCOCO-avg 87.9 (was 57.1 in LFM2-VL-3B) - Multi-image: BLINK 61.5 (was 50.2) · MuirBench 58.3 (was 34.9) - Function calling with vision: ToolSandbox 59.5 (was 26.4) · BFCL v4 32.5 (was 20.5) - Multilingual visual: MMMB 83.0 · Multilingual MMBench 79.5 - STEM visual: MathVista mini 68.5 · MMMU val 48.4 · MMMU-Pro 30.5 · LogicVista 37.4 Text-only instruction following lags the field — IFBench 25.8 / Multi-IF 59.4 trail Gemma-4-E2B (34.1 / 69.4). For pure text, the LFM2.5-2.6B base is the better pick. ## Strengths - Edge-class performance floor. Ties InternVL 3.5 4B at ~65% of the parameters; beats Gemma-4-E2B (5.1B) by a wide margin. - Standout screen / UI understanding. ScreenSpot-v2 jumps from a near-zero 6 / 7.6 / 2.5 on the predecessor to 78 / 81 / 82 — the single biggest UX-relevant gain in the release, and the reason alphasignal.ai picked “screen understanding and grounding” as the win framing. - Real on-device throughput. 228 tok/s on M5 Max, 116 tok/s on Ryzen AI Max+, 20 tok/s on a phone — not paper benchmarks. The phone number is the one that matters for ABS’s privacy story. - Grounding and function-calling leap. RefCOCO 57.1 → 87.9; ToolSandbox 26.4 → 59.5 — turn a VLM into something

Sources

• Liquid AI blog — LFM2.5-VL-3B: A Better and Faster Vision-Language Model for the Edge (Aug 12, 2026)
• Hugging Face model card — LiquidAI/LFM2.5-VL-3B
• Liquid AI Docs — LFM2.5-VL-3B model page
• unite.ai — Liquid AI Ships LFM2.5-VL-3B for Faster Vision-Language AI on the Edge (Aug 12, 2026)
• VentureBeat — No cloud, no GPUs, no problem: Liquid AI’s new model LFM2.5-2.6B (Aug 6, 2026)
• Unsloth — LFM2.5 run + fine-tune guide
• alphasignal.ai — Liquid AI’s LFM2.5-VL-3B beats models twice its size, running on your phone (Aug 12, 2026)