이미지: X — 모델·오픈소스 화면 갈무리
Summary
- Liquid AI released LFM2.5-VL-3B, a lightweight vision-language model that reads screens, documents, and real-world images, on Hugging Face
- At Q4_K_M quantization, using about 3GB of memory, it delivered 228 tokens/sec on Apple's M5 Max and 20 tokens/sec on the Galaxy S26 Ultra
- On a single H100 with vLLM 0.26, the company reported 34ms time-to-first-token for multi-image input and output throughput of about 11,000 tokens/sec under high concurrency
- 모델명
- LFM2.5-VL-3B (파라미터 3.1B)
- 공개일·경로
- 2026년 8월 12일, 허깅페이스 배포
- 온디바이스 메모리
- 약 3GB (Q4_K_M 양자화 + F16 비전 프로젝터)
- 디코드 속도
- Apple M5 Max 228 tok/s · AMD Ryzen AI Max+ 395 116 tok/s · Galaxy S26 Ultra 20 tok/s
- GPU 실측 조건
- NVIDIA H100 SXM5 1장, vLLM 0.26, BF16, 512² 이미지
- GPU 처리량
- 멀티이미지 첫 토큰 34ms, 고동시성 출력 약 11K tok/s, 하루 약 10억 출력 토큰
- 학습 변경점
- 비전 사전학습 토큰 4배 확대, 어휘 사전 12만8000개로 2배 확장
- 벤치마크
- General 평균 70.1, Multilingual 81.2 (자체 집계)
An eye that fits in 3GB
A model that looks at an image and writes an answer on a single Galaxy S26 Ultra spits out 20 tokens per second. No cloud, no account, no upload. That's the measured figure for LFM2.5-VL-3B, the vision-language model Liquid AI unveiled on the 12th. It has 3.1 billion parameters, and at 4-bit quantization its memory footprint is about 3GB. Given that mid-range smartphone RAM these days runs 8-12GB, that means a single model for reading images has shrunk down to a size that can live resident on a phone.
What the model reads
The company positions the use cases in three categories. First, screens — reading mobile, web, and desktop UIs. Second, text and tables in documents and charts. Third, real objects coming in through a camera. Two capabilities are layered on top. One is grounding — the ability to point to a specific object in an image using coordinates. Because it answers with "this coordinate on the screen" rather than "this button," it can serve as the hands of an agent that operates a screen on someone's behalf. The other is tool calling, which the company says works not just from text but also from image input.
Device-by-device benchmarks — alongside rival models
The attached benchmark table was measured with a 512×512 image plus 1,024 text tokens as input, generating 64 tokens. One interesting point: it doesn't always take first place on speed. Looking purely at decode speed on the M5 Max, InternVL3.5-2B (260) and Qwen3.5-2B (237) are faster. Instead, LFM2.5-VL-3B strikes a balance between time-to-first-token and memory usage.
| Apple M5 Max | Time to first token (ms) ↓ | Decode (tok/s) ↑ | Memory (MB) ↓ |
|---|---|---|---|
| LFM2.5-VL-3B | 248 | 228 | 3,257 |
| gemma-4-E2B | 271 | 172 | 4,938 |
| gemma-4-E4B | 441 | 111 | 7,204 |
| InternVL3.5-2B | 438 | 260 | 2,759 |
| Qwen3.5-4B | 406 | 126 | 5,172 |
Things look different on a phone. On a Snapdragon-based Galaxy device (SM-S948U1), time-to-first-token was 13,092ms — 13 seconds. Still, among the models compared, InternVL3.5-4B took 65 seconds. The numbers show that image understanding on a phone is, for now, still a "wait for it" task rather than an "instant answer."
A different face on a server
Using a single H100 SXM5 with vLLM 0.26, the company presented three figures. Time-to-first-token for multi-image input was 34ms — compared against roughly 200ms for the Gemma family. Under conditions with heavy concurrent requests, output throughput reached about 11,000 tokens/sec, roughly double that of the 4B-class models tested. And on a single GPU, the model can produce about 1 billion output tokens per day. The claim is that the value of a small model isn't limited to "running on a phone." Handling more requests on the same GPU also lowers the per-unit cost of document- and screen-processing pipelines.
vLLM adds same-day support for Nemotron 3.5 Lightning, layers on three inference accelerations
How it was trained
The base is the company's own LFM2.5, with this version scaling up the vision side. Vision pretraining tokens were quadrupled, and the data mixed image-caption, OCR, grounding, and instruction-following sets, both curated and synthetic. Rather than building a new tokenizer, the company expanded the existing one, doubling the vocabulary to 128,000 tokens — a factor read as behind its lead in multilingual scores.
| In-house evaluation | General average | Multilingual |
|---|---|---|
| LFM2.5-VL-3B | 7070.1 | 81.2 |
| LFM2-VL-3B (previous version) | 6969.1 | 79.1 |
| gemma-4-E4B-it (8B) | 6363.2 | 75.8 |
| InternVL3.5 4B | 7069.7 | – |
| Qwen3.5-4B | 7069.7 | – |
| gemma-4-E2B-it (5.1B) | 5656.2 | 68.0 |
These are the manufacturer's own figures, so they should be taken with that caveat. Still, the reported gains over the previous version — 1.0 point on General and 2.1 points on Multilingual — indicate the scale of this update.
So what changes
Until now, "AI that looks at a screen and clicks" has generally required uploading a screenshot to a server. Many of those images — internal documents, banking app screens — are uncomfortable to send elsewhere. If a 3GB-class model can run on a laptop or phone, there's room to process those images without ever sending them out. This is the opposite aim from when Naver uploaded HyperCLOVAX-SEED-Think-32B, a 32-billion-parameter-class vision-language model, to Hugging Face on August 9. One is a server model built for deep reasoning; this one is a device model built to answer fast.
The company itself drew a line: it's suited for tasks that need quick responses, such as screen/UI agents, document and chart reading, grounding, and multi-image input, while problems that require step-by-step reasoning should be handled by tuning it through a general post-training pipeline. The model is available for direct download from Hugging Face.
Naver unveils SEED-Think, a 32B-class reasoning vision-language model
