Review of DeepSeek V4 Pro
After 3 months of daily use, I have enough to say. DeepSeek V4 Pro is the first DeepSeek model that I actually use in production. Here is what works, what does not, and whether it is worth your money.
The headline: 128K context window. Most open-weight models cap at 8K to 32K. DeepSeek V4 Pro handles 128K tokens without losing coherence. I tested this by feeding it a 60-page technical specification and asking it to summarize specific sections. The model tracked details across the entire document. The closed-frontier models still do this better, but not by much. The catch: 128K context is slow. First-token latency is around 2 seconds on the Pro variant. The smaller variants are faster but less accurate on long documents.
My use case for V4 Pro is mostly code review and documentation generation. The model's coding ability is genuinely good. I tested it on 50 PRs in my saas.pet repo. The review quality was on par with GPT-4 for routine code. For complex refactors, it still loses to Claude. The pricing makes up for the gap. I pay $0.14 per million input tokens for V4 Pro. The closed alternatives cost 5-10x more. For a solo founder with a $200/month AI budget, V4 Pro is the obvious choice.
The workflow: I take every PR that comes in, paste it into V4 Pro, and ask for a review. The review takes 30 seconds. I edit the review for tone and accuracy. I post the review on the PR. This used to take 15 minutes per PR. With V4 Pro, it takes 2 minutes. The catch: the model is not as good at spotting subtle architectural issues as the closed alternatives. I have found 3 cases where it missed a problem that GPT-4 caught. For routine code, it is reliable. For architecture reviews, I still use Claude.
V4 Pro API pricing: $0.14 per million input tokens, $0.28 per million output tokens. The smaller V4 Mini: $0.03 per million input, $0.06 per million output. The smallest V4 Nano: $0.005 per million input, $0.01 per million output. For comparison: GPT-4 is $5 per million input. Claude Sonnet is $3 per million input. V4 Pro is 30x cheaper than GPT-4 for similar quality on most tasks.
Self-hosting is an option for those with the hardware. The 7B parameter variant runs on a single A100. The 70B variant needs 4x A100s. The Pro variant (the largest, ~400B parameters) needs significant infrastructure. I do not self-host. The API pricing is good enough. The catch: rate limits. Free tier is 5 RPM. Pro tier is 60 RPM. For batch processing, that is not enough. For interactive use, it is plenty.
V4 Pro's small variants (Mini and Nano) are nearly as good as the full Pro for most tasks. The benchmark scores show a 10-15% gap. In real use, the gap is much smaller. For 80% of my tasks, Mini is good enough. I only use Pro for the most complex reasoning tasks. The implication: you can save a lot of money by using the right model for the task. Don't default to Pro. Default to Mini. Upgrade to Pro when the task demands it.
The other thing nobody tells you: the open-weight versions are not as good as the hosted API. DeepSeek releases the weights, but the inference quality is lower than the API. The API has been optimized with custom routing. The weights are the raw model. For production, use the API. For research, use the weights. Mixing the two gives you the worst of both worlds.
It is 70-80% as good as GPT-4 for most tasks. For coding, it is 85% as good. For reasoning, it is 75% as good. For creative writing, it is 50% as good. The cost is 30x lower. For most production use cases, 70-80% at 30x lower cost is a better deal than 100% at 30x higher cost. The gap will close as DeepSeek iterates. V5 is rumored to be coming in Q4 2026.
The Mini variant (7B) runs on 16GB of unified memory. The Nano variant runs on 8GB. The Pro variant does not. Pro needs 80GB+ of VRAM. If you are a solo dev, Mini is what you run locally. The full Pro is API only. The quality difference between Mini and Pro is smaller than you would expect. For most tasks, Mini is good enough. Save your Pro budget for the tasks that matter.
The weights are released. The training data is partially released. The API is proprietary (in the sense that you cannot run it on your own hardware at full quality). The open-weight versions are usable but lower quality than the API. If open source matters to you, V4 is more open than closed alternatives. If quality matters, the API is the way to go. The choice is yours.
