DeepSeek V4-Flash: Open-Weight, Million-Token Context
DeepSeek's V4 preview, released April 24, made a 1M-token context window the default for an open-weight model family, under an MIT license. Three weeks in, the more interesting model of the pair is not the 1.6T-parameter flagship. It is DeepSeek V4-Flash: 284B total parameters, 13B active, and benchmark results that land within a couple of points of the Pro model at a fraction of the serving cost.
The thesis is simple. V4-Flash is the first open model where "1M context" is cheap enough to use routinely, not just a spec-sheet number. At launch it cost $0.14 per million input tokens and $0.28 per million output through DeepSeek's API.
What DeepSeek released
- V4-Flash: 284B total / 13B active parameters, 1M context, FP4 experts with FP8 elsewhere, MIT license.
- V4-Pro: 1.6T total / 49B active parameters, 1M context, MIT license.
Both are mixture-of-experts models with a new hybrid attention scheme that combines token-wise compression with DeepSeek Sparse Attention. DeepSeek's model card says that at 1M tokens, V4-Pro needs 27% of the per-token inference FLOPs and 10% of the KV cache that V3.2 needed. The KV cache figure is what makes long context affordable: memory, not compute, is usually what limits serving at that length.
The benchmark picture
These are DeepSeek's own numbers in the highest reasoning mode ("Max"), from the model card:
| Benchmark | V4-Flash Max | V4-Pro Max |
|---|---|---|
| LiveCodeBench | 91.6 | 93.5 |
| Codeforces rating | 3052 | 3206 |
| SWE-bench Verified | 79.0 | 80.6 |
| MMLU-Pro | 86.2 | 87.5 |
Flash gives up 1.3 to 1.9 points on the headline coding and knowledge benchmarks while activating about a quarter of Pro's parameters per token. For Pro, DeepSeek also reports 83.5 on MRCR at 1M tokens and 62.0 on CorpusQA at 1M, which suggests the long window holds up for retrieval rather than collapsing past a few hundred thousand tokens. Treat all of these as vendor-reported until third-party runs confirm them.
Where it fits
V4-Flash is not the only 1M-context option. Grok 4.3 ($1.25/$2.50), Gemini 3.1 Pro Preview ($2/$12), and Claude Opus 4.7 ($5/$25) all offer it too. The difference is the combination:
- Price: at $0.14/$0.28, V4-Flash is roughly 9x cheaper on input than Grok 4.3 and 14x cheaper than Gemini 3.1 Pro on input.
- Open weights: you can self-host, fine-tune, and inspect activations. None of the closed options allow that.
- Hardware reality: 13B active parameters keeps per-token compute low, but all 284B parameters still have to sit in memory. In FP4 that is roughly 150GB of weights before the KV cache. This is a multi-GPU server model, not a laptop model.
What to watch
- Independent evals. The Max-mode numbers use DeepSeek's own harness. Look for third-party SWE-bench and long-context retrieval results before you replace a production model.
- Preview pricing. Preview prices rarely survive general availability unchanged. Budget for them to move.
- Serving support. The new attention scheme needs inference-engine support to deliver its memory savings when you self-host. Check that your stack supports it before you plan on self-hosting at 1M tokens.
Update, September 2026: the preview price did not last. DeepSeek moved V4 to peak and off-peak billing on August 16; V4-Flash went to $0.22/$0.66 off-peak and double that at peak. See DeepSeek Peak Pricing for the details.
Takeaway
If you need long context and your data can go to a Chinese API provider, V4-Flash is the cheapest serious option available today. If it can't, the MIT weights mean you can run it yourself, but plan on server-class hardware. Either way, benchmark it on your own long documents before trusting the model card.
Related: Longest-Context Models · Best Open-Weight Models