What it is
DeepSeek-V4-Pro is the large model in DeepSeek's V4 family, first published on Hugging Face in April 2026 and still labelled a preview on its model card. DeepSeek's API now serves an August 13 update, DeepSeek-V4-Pro-0813; the Hugging Face card, last changed in June, does not mention it, so it is unclear whether those updated weights are public. It is a mixture-of-experts model with 1.6 trillion parameters, 49 billion active per token, and a 1M-token context. It handles text only.
It has three modes: Non-think for fast answers, Think High for careful reasoning, and Think Max, which uses a special system prompt and needs a context window of at least 384K tokens. The weights ship as mixed FP4 and FP8 under the MIT licence.
On DeepSeek's API it costs $1.32 per million input tokens (cache miss) and $3.96 per million output tokens at peak hours, and half that off-peak. Cached input costs $0.022 to $0.044. Output can run up to 384K tokens.
Our review
V4-Pro is a frontier-class model you can download and use for anything, and also a spring 2026 model being sold in autumn. DeepSeek's card compares Think Max against Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro. It led that group on LiveCodeBench (93.5) and Codeforces (3206) and tied Gemini on SWE-bench Verified at 80.6, a fraction behind Opus 4.6's 80.8. These are DeepSeek's own figures, and every rival named has since been replaced.
The weaknesses are consistent: knowledge and agentic tool use. It scored 37.7 on Humanity's Last Exam against Gemini's 44.4, and 48.2 on the tool-assisted version, last of the six models in the table. On a 1M-token retrieval test (MRCR) it reached 83.5 against Opus's 92.9, so the long context is real but less reliable at the far end.
Price is where it wins. Off-peak, which DeepSeek defines as everything outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, input costs $0.66 and output $1.98 per million. That is a sixth and a tenth of Opus 5.5's $4 and $20. The catch is a concurrency limit of 500 requests per account on Pro, against 2,500 on DeepSeek's own Flash model, which is also cheaper and the only one of the two that reads images.
The hosted API and the open weights are different products for compliance purposes. DeepSeek's privacy policy says it stores data in the People's Republic of China and lists training on user input among its purposes, with opt-out by email. Self-hosting removes that, but the checkpoint is about 865 GB, so it needs a full eight-GPU H200 or B200 server before any memory is left for a long context.
Where it earns its place
A platform team at a European bank with its own GPU cluster
Run a frontier-class coding assistant on-premises where no data can leave the building
Serve the MIT-licensed weights with vLLM or SGLang on an eight-GPU H200 or B200 node. The licence permits this with no fee, but the hardware is the bill: the checkpoint alone is about 865 GB.
A two-person startup building a batch code-review bot for open-source repositories
Run millions of tokens of reasoning a day at the lowest possible price
Schedule jobs off-peak on DeepSeek's API and keep repository context cached. A billion input tokens a month at 70% cache hits plus 100M output costs about $410 off-peak, against roughly $3,340 on Opus 5.5.
A competitive-programming training platform
Generate and verify solutions to contest-style problems
Use Think Max, where V4-Pro posts its best scores (LiveCodeBench 93.5). Think Max needs at least 384K of context, so long runs use more tokens and more of the 500-request concurrency cap.
Before you commit
- Setup
- On the API, select deepseek-v4-pro; thinking is on by default and can be switched off. For self-hosting, the model card gives vLLM, SGLang, Docker and Transformers commands and recommends temperature 1.0, top_p 1.0.
- Integrations
- DeepSeek API (also usable from Claude Code and Codex, where search and fetch calls are free until November 1, 2026), Hugging Face weights, vLLM and SGLang.
- Data and privacy
- Open weights are MIT-licensed with no commercial or revenue restrictions. DeepSeek's privacy policy (February 2026) says data is stored in China, lists model training among its uses of user input, and offers opt-out only by emailing privacy@deepseek.com.
- Cost at scale
- On the API, 1B input tokens a month at 70% cache hits plus 100M output tokens costs about $820 at peak rates or $410 off-peak. Self-hosting needs roughly 865 GB for weights alone, which by our arithmetic means at least an eight-GPU H200 server (1.1 TB of HBM) and more for full 1M-token contexts.
- Limits
- 1M context, 384K output, 500 concurrent requests per account. No image input (Flash only). Trails its comparison group on knowledge (HLE 37.7) and is weaker than Opus 4.6 at the far end of the 1M context.
What's good
- MIT licence on full weights, no revenue caps or use restrictions
- $0.66 / $1.98 per million off-peak, cache hits from $0.022
- Top competitive-programming scores in its own comparison (Codeforces 3206)
- 1M-token context, 384K output
What's not
- Weakest of its comparison group on tool-assisted HLE (48.2)
- Hosted API stores data in China; policy allows training on inputs
- About 865 GB of weights; 500-request concurrency cap on the API
- Text only
Best for
Teams that want frontier-class open weights for coding and reasoning, either on their own eight-GPU server or through a very cheap API for non-sensitive work.
Alternatives
DeepSeek-V4-Pro: common questions
- Is DeepSeek-V4-Pro worth it?
- DeepSeek-V4-Pro is frontier-class reasoning under an MIT licence at about a quarter of Opus 5.5's per-token price, provided your data can go to China or you own a rack of H200s. We rate it 4.3 out of 5.
- How much does DeepSeek-V4-Pro cost?
- Free MIT-licensed weights. DeepSeek API: $0.66 input / $1.98 output per 1M tokens off-peak, $1.32 / $3.96 at peak; cache hits from $0.022.
- What are the downsides of DeepSeek-V4-Pro?
- Weakest of its comparison group on tool-assisted HLE (48.2); hosted API stores data in China; policy allows training on inputs; about 865 GB of weights; 500-request concurrency cap on the API; text only.
- Who is DeepSeek-V4-Pro best for?
- Teams that want frontier-class open weights for coding and reasoning, either on their own eight-GPU server or through a very cheap API for non-sensitive work.
- How much does DeepSeek-V4-Pro cost for a team or in production?
- On the API, 1B input tokens a month at 70% cache hits plus 100M output tokens costs about $820 at peak rates or $410 off-peak. Self-hosting needs roughly 865 GB for weights alone, which by our arithmetic means at least an eight-GPU H200 server (1.1 TB of HBM) and more for full 1M-token contexts.
- Is DeepSeek-V4-Pro safe for company data?
- Open weights are MIT-licensed with no commercial or revenue restrictions. DeepSeek's privacy policy (February 2026) says data is stored in China, lists model training among its uses of user input, and offers opt-out only by emailing privacy@deepseek.com.
- What are the best DeepSeek-V4-Pro alternatives?
- The closest alternatives we have reviewed are DeepSeek-V4.1-Flash, Kimi K3, GLM-5.3 and Qwen3.8-Max.
Answers checked Oct 22, 2026.
Make DeepSeek-V4-Pro? Show this rating on your site
