The bitter-lesson fix for MLLMs: whole image at native resolution, up to 16k tokens
gabriberton · x · 2026-10-19
Part 5 of gabriberton's MLLM thread: the bitter-lesson solution most labs moved to is simply feeding the whole image to the model at native resolution, up to a high token cap (e.g. 16k tokens in Qwen3-VL). Examples: Qwen3-VL, Kimi-VL, Gemma 4 — avoiding the per-model crop-count/resolution/overlap hyperparameter tuning of multi-crop schemes.
Related event: Four Years of MLLM Visual Encoding: From Cropping to Native Multimodal(8 posts)→
More from Models
- Timezone conversion was my last use case for Google — and LLMs now handle it — shauseth · 2026-10-19
- DeepSeek's V4.1-Flash-1017: 500M tokens for ~$4, sparking heavy testing — sven_ai · 2026-10-19
- Qwen Image 2.1's built-in prompt enhancer is unusable: 12k tokens still not enough — dtdisapointingresult · 2026-10-19
- 93-Model Benchmark Over 1.5 Years: Distillation + Quantization + RL Hurts General AI — Afinetheorem · 2026-10-19
- Dev comparison: Opus shines at UI, but Astra still better at logic and completeness — mallow610 · 2026-10-19
- DeepSeek's V4.1 Flash called a stopgap: V4.1 Pro, DSec and infra reportedly in the works — teortaxesTex · 2026-10-19