DeepSeek Releases V4 Flash Vision Weights Under MIT — 305B Parameters, 168 GB, Free to Self-Host
DeepSeek published open weights for V4 Flash Vision Exp on August 31 under an MIT license, making the full checkpoint available for local deployment. The model opened as an API on August 21. The weights release — disclosed initially through a ModelScope post on X — is a separate event: developers can now download the full checkpoint and serve it on their own infrastructure without DeepSeek’s API in the path.
Scale and Format
The Hugging Face repository lists DeepSeek-V4-Flash-Vision-Exp at 305 billion total parameters, distributed across 48 Safetensors files totalling 168 GB. The release includes a tokenizer, prompt-encoding reference and a minimal PyTorch inference implementation. Full vLLM and SGLang integration is the expected deployment path for production use.
The MIT license is the broadest DeepSeek has used for this family. It allows modification, fine-tuning and incorporation into commercial products, provided downstream builders can supply the hardware.
What the Benchmarks Show
DeepSeek’s model card publishes multimodal agent evaluation results against Claude Opus 4.8:
| Benchmark | V4 Flash Vision | Opus 4.8 |
|---|---|---|
| Agent’s Last Exam | 27.3 | 25.7 |
| ZeroBench (pass@5) | 35.0 | 34.0 |
| Chartography | 64.3 | 65.0 |
| ApexBench | 36.5 | 39.4 |
The Vision model leads on two of the four multimodal agent evaluations and trails on two. The stronger numbers are on tasks that combine web navigation or screenshot interpretation with multi-step reasoning; Chartography and ApexBench favor Opus 4.8’s existing multimodal pipeline.
Across seven text-agent tests in the model card, V4 Flash Vision reports one win over Opus 4.8 — on DeepSWE — and six losses. DeepSeek describes these results accurately as narrowing gaps rather than leading scores.
Why This Matters
The API-first then weights release pattern lets DeepSeek gather usage data and identify issues before exposing the checkpoint to the broader community. The 10-day gap between API launch and weights release is shorter than the V4 Pro cycle, which saw a preview period of several months before weights shipped.
For practitioners, the MIT license means V4 Flash Vision is now available as a base for fine-tuning specialized visual agents: document processing, screenshot-driven automation, chart interpretation. At 305B parameters, it requires serious GPU hardware — a single H100 node at 80GB vRAM cannot load the full BF16 checkpoint — but multi-GPU setups with NVLink or high-bandwidth interconnects are the standard deployment environment for models in this size class.
The commercial import is most significant for teams that cannot use external APIs due to data residency, latency or contractual constraints. DeepSeek’s V4 Flash family is already priced aggressively on its own API; having an MIT-licensed checkpoint means enterprise buyers with on-premises infrastructure can run it without any per-token cost.
DeepSeek’s own view, stated in the model card, is that V4 Flash Vision closes but does not erase the gap with frontier proprietary models on multimodal agent tasks. That framing is unusually honest. The benchmark table supports it.