DeepSeek V4 review: Pro vs Flash, a 1.6T MoE built for long context
DeepSeek V4 (deepseek-v4-pro and deepseek-v4-flash) is a strong pick when you need to push a lot of very long context, up to around 1M tokens, through a model at a cost you can live with. Its architecture was designed for exactly that, and it is the reason to choose it.
Specs: a 1.6T-parameter MoE
DeepSeek released V4 on April 24, 2026 as a family of open-weight Mixture-of-Experts models. The larger V4 Pro has 1.6 trillion total parameters, of which only about 49B are active for each token.
- Context: about 1M tokens
- Max output: about 384K tokens (Pro)
- Input and output: text only
- License: MIT
Both deepseek-v4-pro and the lighter deepseek-v4-flash are on FastMetal.
Architecture: cutting the cost of long context
The core change in V4 is a rebuilt attention mechanism. It combines CSA (Compressed Sparse Attention) and HCA (Heavily Compressed Attention), and at 1M tokens of context DeepSeek reports per-token inference FLOPs at 27% of V3.2's and the KV cache at 10%.
That is the gap between a model that accepts a long context and one you can afford to run at long context. A spec sheet's context length hides it; your bill does not.
Benchmarks, read carefully
On coding, the top variant of the series, V4-Pro-Max, was reported at 80.6% on SWE-bench Verified, the best open-weight result at release and level with Gemini 3.1 Pro.
That number belongs to the Max variant, which is a different model from the deepseek-v4-pro FastMetal serves. Treat it as a guide to what the family can do, and measure the model you call on your own work.
Cost
Price-performance is V4's selling point: at launch it was reported as much cheaper per output token than the closed frontier models. Both models are billed per million input and output tokens; current prices are on the DeepSeek V4 Pro and DeepSeek V4 Flash model pages.
On long-context batch work that difference shows up directly in the monthly total. The bill is still price times volume, so set a prepaid limit before you scale up.
Japanese and other languages
V4 is multilingual and gives usable answers on general Japanese tasks. We found no Japanese-specific benchmark for it: DeepSeek's published results center on math and coding. If writing quality in a particular language decides your output, compare it on your own prompts first.
Good fit, poor fit
Good fit
- Batch jobs over contexts approaching 1M tokens
- Work where cost comes first but reasoning still needs to be solid
- Math and coding
- Commercial projects that want a permissive license
Poor fit
- Multimodal work with images or audio (V4 is text only)
- Output where fine-grained Japanese style decides the quality
- Expecting the headline benchmark score (that is the Max variant)
FAQ
Q. How do I choose between Pro and Flash?
Use deepseek-v4-pro when quality comes first and deepseek-v4-flash when speed and cost do. The same key calls both, so the reliable answer is to run your real prompts through each.
Q. Can I use DeepSeek V4 commercially?
Yes. It is released under the MIT license.
Q. How is it in Japanese?
Fine for general use, but we found no Japanese-specific benchmark. Test it on real prompts before relying on it for anything important.
For head-to-head runs, see DeepSeek V4 Pro vs Claude Opus 5 and DeepSeek V4 vs Claude Fable 5.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-pro",
"messages": [{"role": "user", "content": "Hello"}]
}'
Change model to deepseek-v4-flash to compare it with the lighter version. Every model is in the model catalog.