GLM-5.3-Flash lidera los modelos de visión gratuitos de FreeLLMArena con una puntuación de 1272
GLM-5.3-Flash lidera los modelos de visión gratuitos en 1272, cabiendo en Mac Studio M3 Ultra 512 GB; Claude Fable 5 lidera la lista de pagos con 1309.
FreeLLMArena ranks many open-weight vision models. The highest-scoring free model is GLM-5.3-Flash with a score of 1272, behind the top paid model Claude Fable 5 (score 1309).
How the ranking works
FreeLLMArena collects blind human votes from the Arena platform. Each model receives a score that reflects usefulness, accuracy, and overall user satisfaction. The list below includes only models that have open weights.
Top free vision models
| Puesto | Modelo | Laboratorio | Puntuación | Parameters* | Approx. 4-bit size | Fits fully on |
|---|---|---|---|---|---|---|
| 36 | GLM-5.3-Flash | zai | 1272 | 321.3 B | ~197 GB | Mac Studio M3 Ultra 512 GB or bigger |
| 39 | Gemma 4 31B | 1265 | 31.3 B | ~22 GB | RTX 3090 or bigger | |
| 40 | Kimi K2.6 | moonshot | 1265 | 1026.9 B | ~625 GB | no single home machine |
| 42 | DeepSeek-V4.1-Flash-20260910 (Max) | deepseek | 1263 | size not public | n/a | n/a |
| 51 | Kimi K2.5 (Thinking) | moonshot | 1252 | 1026.9 B | ~625 GB | no single home machine |
| 56 | MiMo-V2.6-Flash | xiaomi | 1248 | 310.8 B | ~191 GB | Mac Studio M3 Ultra 512 GB or bigger |
| 58 | MiMo-V2.6-Pro | xiaomi | 1247 | 1024.2 B | ~623 GB | no single home machine |
| 59 | Qwen3.5-397B-A17B | alibaba | 1246 | 403.4 B | ~247 GB | Mac Studio M3 Ultra 512 GB or bigger |
| 61 | Qwen3.8-27B | alibaba | 1243 | 27.8 B | ~20 GB | RX 7900 XT or bigger |
| 63 | Gemma 4 26B A4B | 1242 | 25.8 B | ~18 GB | RX 7900 XT or bigger | |
| 67 | Kimi K2.5 (Instant) | moonshot | 1236 | 1026.9 B | ~625 GB | no single home machine |
| 70 | MiMo-V2.5 | xiaomi | 1234 | 310.8 B | ~191 GB | Mac Studio M3 Ultra 512 GB or bigger |
*Parameters refer to the number of trainable weights. “B” stands for billion. The 4-bit size is the storage needed when the model is quantized to 4-bit precision.
Choosing a model for your hardware
- Memory requirement: Use the “Fits fully on” column to see the minimum hardware that can hold the 4-bit model.
- Model size: Larger parameter counts need more memory and may need multi-GPU or cloud resources.
- Task fit: Select a model whose size matches your performance and resource constraints.
Where to find the models
All listed models have open weights available from their respective labs. FreeLLMArena provides links to the repositories and benchmark results in the Vision category. Visit the Vision page to explore the full ranking and download the models.
¿Es Kimi K3 (Max) el modelo adecuado para ti?
Consulta todos los modelos gratuitos clasificados por votos humanos ciegos, con tamaño, memoria y velocidad en una GPU doméstica.