DeepSeek-V4-Flash-Vision-Exp
The V4 arc completes in threes: Flash, Pro, and now Vision — all under the same 1.08 kB MIT file
What it is
DeepSeek's first multimodal model in the V4 family: the V4-Flash architecture with native visual modules incorporated and continued training to unlock visual understanding — not a bolted-on vision tower. Ten days as an API-only experiment, then the weights landed August 31 with an unusually complete repo: tokenizer, prompt-encoding reference, and a minimal PyTorch inference implementation covering the vision encoder and aligner, DFlash attention, MoE, Hyper-Connections, and the DSpark speculative-decoding path — draft and target weights in one checkpoint.
Why it's interesting
This sat on our watch list as 'API-only, watch for weights' — and the weights came, with the license the V4 line has now made policy: the LICENSE file is the standard, unmodified 1.08 kB MIT text, the card states 'This repository is licensed under the MIT License,' and the weight shards live in the same repo under it. The multimodal jump over text-only Flash is real on DeepSeek's own table — 36.5 vs 26.2 on ApexBench, 35.0 ZeroBench Pass@5 (a point past Opus 4.8) — while text-agent scores hold steady (83.9 Terminal-Bench 2.1). Third MIT release in the V4 family in a month; at this point the clean license is the least surprising thing about a DeepSeek drop.
Use cases
- Screenshot-driven and UI-reading agent loops, self-hosted — the multimodal benchmarks are the pitch
- Vision-capable coding and terminal agents that keep Flash-class text performance
- Studying a lab-grade multimodal reference implementation — encoder, aligner, and DSpark path in readable PyTorch
Who it's for
Agent builders who wanted V4-Flash with eyes; infra teams with a four-GPU node
Setup
Advanced. Reference serving is a 4-GPU node (vLLM recipe targets 4×GB300 with FP8 KV cache; SGLang with --speculative-algorithm DSPARK); the repo's minimal PyTorch implementation covers conversion and single-prompt inference
Limitations & cautions
'Exp' is in the name on purpose — DeepSeek calls it experimental and the model ID may not survive evaluation; the benchmark table is self-reported, evaluated with DeepSeek's own Harness in minimal mode; the card doesn't restate parameter counts or context length (the API lists 1M context and ~384 tokens per image — confirm before you plan around either); and there's no document/OCR mode, so PDFs mean rasterizing first.
Editorial takeaway
Flash was the promise, Pro was the proof, and Vision is the habit. When a lab's third straight release in a month ships the same one-kilobyte MIT file, you stop calling it a license choice and start calling it an identity.