LFM2.5-VL-450M
Liquid AI · 2026-04-08 · 449M parameters
LFM2.5-VL-450M is Liquid AI's April 8, 2026 refresh of LFM2-VL-450M on the LFM2.5-350M backbone, with an 86M-parameter shape-optimized SigLIP2 NaFlex vision encoder that tiles large images into 512x512 patches plus a thumbnail. The card gives a 32,768-token context and nine languages. The BF16 checkpoint contains 448,718,848 parameters, and the tokenizer assigns 64,400 distinct IDs within 65,536 embedding rows.
Benchmark scores
| Benchmark | Score |
|---|---|
| BFCL v4 Overall Accuracy | 21.1 |
| BLINK | 43.9 |
| Capability | 88.8 |
| CountBenchQA | 73.3 |
| IFEval - release-reported, metric unspecified | 61.2 |
| InfographicVQA (val) | 43.0 |
| MM-IFEval | 45.0 |
| MM-Vet | 41.1 |
| MMLU Pro | 19.3 |
| MMMU | 32.7 |
| MMStar | 43.0 |
| Multi-IF | 34.6 |
| OCRBench | 684 |
| POPE (F1, pooled over three settings) | 86.9 |
| RealWorldQA - release-reported visual perception | 58.4 |
| RefCOCO-M - release-reported, metric unspecified | 81.3 |