DeepSeek-V4-Flash
DeepSeek · China · MIT
Commercial use: Yes — free for commercial use
MIT. Full commercial use, redistribution and distillation of outputs, with no user threshold and no attribution requirement. The most permissive licence available on a model at this capability level.
A reasoning model as capable as the paid frontier options, with the weights published under MIT. You can download it and run it yourself.
Why it should matter to you
If your blocker on AI has been 'we cannot send this data to a third party', this is the class of model that removes the blocker rather than working around it. The trade is real: you need serious hardware, and the reasoning traces are long and slow. But there is no vendor in the data path and no per-token bill.
How it connects to our work
The distinction we spend most time explaining in security reviews: using DeepSeek's hosted API sends your data to Chinese infrastructure and is a legitimate procurement objection. Running DeepSeek's weights inside your own network sends nothing anywhere, and the MIT licence is more permissive than Meta's or Google's. Reviewers frequently block both on the strength of the first.
From our field notesSelf-hosting decouples 'which jurisdiction' from 'which vendor'.
Advantages
- MIT on a frontier-class model, which remains rare and is the reason it dominates open-weight downloads
- The most-downloaded open frontier model on the Hub by a wide margin (4.5M+ on the flagship repo alone)
- A Vision-Exp variant extends the same licence to multimodal work
- Distillation of its outputs is explicitly permitted, so smaller task models can be built from it legally
Disadvantages
- Too large to self-host outside a serious multi-GPU rig or an Ultra-class Mac
- DeepSeek's hosted API runs on Chinese infrastructure, which is a live procurement question in some sectors
- Reasoning-style outputs are verbose and token-expensive if you do not cap them
Reach for it when
Frontier-quality work where the licence has to survive legal review, and as the teacher model when distilling a small in-house model you intend to own.
Where it falls down
Anything latency-sensitive on your own hardware, and any deployment where a security reviewer will not separate the hosted API from the weights. For local work, use a distilled or smaller model instead.
Running it
Hosted, or 256 GB+ unified memory at Q4. Not a laptop model. See the hardware sizing tables for how that maps to specific chips and cards, and the quantisation guide for what you give up at each bit width.
Free tiers carrying this model
| Provider | Type | The catch | Live limits |
|---|---|---|---|
| OpenRouter | Aggregator | Free-pool membership changes without notice — a model you built on can stop being free, and the list above will drift. Data passes through OpenRouter and then the upstream provider, so check the privacy and routing settings before sending anything sensitive. | check → |
| NVIDIA NIM | Inference host | It is a credit grant, not a permanent free tier. When the credits run out you are on a paid plan or self-hosting the container. | check → |
Jurisdiction: China
This is the crucial split most write-ups miss: using a Chinese lab's HOSTED API sends your data to Chinese infrastructure and is a real procurement question. Downloading their OPEN WEIGHTS and running them on your own hardware, or on a Western host, sends nothing anywhere. DeepSeek and Z.ai publish under MIT; Alibaba publishes most of Qwen3 under Apache 2.0. Those are among the most permissive licences on this page.
Watch for: Several US states and a number of government bodies restrict Chinese-hosted AI services on official devices. That restriction is about the hosted service, not about the weights running in your own VPC — but expect to have to explain the difference to a security reviewer.
This model has a published technical report: arXiv:2412.19437 The V3 technical report documents the architecture this line is built on. It is indexed in our research library alongside the work it builds on.
Frequently asked
Can I use DeepSeek-V4-Flash commercially?
MIT. Full commercial use, redistribution and distillation of outputs, with no user threshold and no attribution requirement. The most permissive licence available on a model at this capability level.
What hardware do I need to run DeepSeek-V4-Flash?
Hosted, or 256 GB+ unified memory at Q4. Not a laptop model. Weights alone are roughly 210 GB at Q4_K_M and 400 GB at Q8_0. Add KV cache on top of that, which grows with your context length.
What licence is DeepSeek-V4-Flash released under?
MIT. Full commercial use. Shortest and least restrictive of the common licences; no explicit patent grant.
Where can I use DeepSeek-V4-Flash for free?
Free tiers carrying it include OpenRouter, NVIDIA NIM. Limits differ per provider and change often, so check each provider's own limits page. You can also self-host the weights, which has no rate limit at all.
Similar models
GLM-5.3-Flash
80-96 GB at Q4_K_M. A workstation or Ultra-class Mac.
Qwen3.8 27B
24 GB GPU at Q4_K_M, or a 32 GB Mac at Q6_K.
Mistral Large 3 675B
Multi-GPU datacentre, or hosted on La Plateforme.
gpt-oss-120b
80 GB GPU, or 96 GB+ unified memory.
Wiring DeepSeek-V4-Flash into something real?
We build the evaluation harness, the failover and the cost ceilings around a model like this, so it survives contact with production.
Maps to AI agent development, AI automation development and AI strategy and consulting. Or see it working: our case studies.