DeepSeek gives V4 Flash cheap vision — any image caps at 384 tokens
The experimental model reads screenshots, charts and docs at Flash rates, 1M context — as LiteLLM, SGLang and Ray quietly upgrade the serving stack.
Copy markdown
Cheap vision, finally on the Flash line
DeepSeek-V4-Flash-Vision-Exp adds image input to the Flash model at $0.22/$0.66 per 1M tokens — halved in off-peak hours — while matching plain V4-Flash on text and reasoning. You can now point an agent at screenshots and charts without paying frontier-vision rates.
Every image bills the same — up to 384 tokens
Each image is capped at 384 input tokens regardless of resolution, charged at the normal input rate with no vision surcharge; a 5000x5000 screenshot costs the same as a thumbnail. Over a 1M-token context, that makes a weekend doc- or dashboard-triage agent genuinely cheap to run.
The catches: experimental, API-only, self-scored
The "-Exp" tag means no stability or deprecation guarantees, there are no open weights, and the near-Opus numbers are DeepSeek's own harness — it trails Opus 4.8 on 8 of 11 benchmarks. There's no video, PDF, OCR or grounding mode; it takes text and images in, returns text out.
Elsewhere: the self-hosting serving stack leveled up
LiteLLM v1.98 adds provisioned-throughput billing (flat rates for reserved capacity), shadow evals for its auto-router and per-key caching controls. SGLang's latest release reportedly cuts cold starts ~2.4x (a Qwen3-32B server booting in ~36s vs ~85s on an H100), and Ray Serve adds KV-cache-aware routing that tokenizes at ingress to skip duplicate work.