DeepSeek V4 Open Weights: MIT License (August 2026)
The Short Answer
In August 2026, DeepSeek released the weights for its V4/0731 model under an MIT license — a 284B-parameter mixture-of-experts model with just 13B active parameters. It’s one of the most permissive frontier open-weight releases of the year: commercial use, self-hosting, and redistribution all allowed.
Key Facts
| DeepSeek V4/0731 | |
|---|---|
| License | MIT (fully permissive) |
| Total parameters | ~284B (mixture-of-experts) |
| Active parameters | ~13B per token |
| API price (V4 Pro, off-peak) | $0.435 / $0.87 per MTok |
| API price (V4 Flash, off-peak) | $0.14 / $0.28 per MTok |
| Cache-hit (V4 Pro) | ~$0.043 |
What the MIT License Changes
Many “open” models ship under custom licenses that restrict commercial use above a user threshold or forbid training competing models. MIT does none of that — it’s a permissive license allowing commercial deployment, modification, and redistribution with essentially just an attribution requirement. For teams that want a frontier-class model they can self-host without legal friction, this is the headline.
The 284B / 13B Architecture
V4/0731 is a mixture-of-experts (MoE) design: 284B total parameters but only ~13B active per token. That means inference cost and speed track the 13B active count, not the full 284B — so it runs far cheaper than a dense 284B model while retaining the capacity to match much larger models on many tasks. It’s the same efficiency pattern that made earlier DeepSeek releases cost-competitive.
API vs Self-Host
- API (managed): V4 Pro ~$0.435/$0.87 per MTok off-peak (2× during peak hours 1-4 and 6-10 UTC); V4 Flash $0.14/$0.28. Note the
deepseek-chat/-reasonernames were retired July 24, 2026. - Self-host (MIT weights): you pay only your own GPU/compute — attractive for data-sovereignty needs or very high volume where API markup dominates.
Who Should Care
- Self-hosters wanting a permissively-licensed frontier model.
- High-volume API users comparing DeepSeek’s cheap off-peak rates against Western frontier pricing.
- Enterprises with data-residency requirements that rule out managed APIs.
The Reality Check
An MIT-licensed 284B MoE is a big deal for the open ecosystem — but “open weights” isn’t “cheap to run.” You still need multi-GPU hardware to serve 284B params, even with 13B active. For most teams the off-peak API is the practical path; self-hosting pays off only at serious scale or under strict data-sovereignty rules.
Sources
- DeepSeek — model releases & pricing: api-docs.deepseek.com
- DeepSeek — Hugging Face weights: huggingface.co/deepseek-ai