DeepSeek V4 Flash: The one you actually deploy
284B total, 13B active, a million tokens of context, and a published serving recipe somebody else already tuned. The cheapest active-parameter count on this page.
The architecture
how it is builtThe same attention as Pro
V4 Flash inherits the hybrid CSA and HCA design and the 1,048,576-token context window. It is not a distilled or truncated Pro, it is the same architecture at a smaller expert count.
13B active
The lowest active-parameter count of any model here, at full million-token context. This is the number that makes agent workloads, where a thousand sessions share a long system prompt, affordable rather than theoretical.
Mixed FP4 and FP8
Same checkpoint strategy as Pro: MoE expert parameters in FP4, most everything else in FP8. DeepSeek publishes the format but not a byte count, so the 150 GB figure below is derived and flagged as such wherever it appears. Worth noting the loaded model reports closer to 291B than the headline 284B, which is the sort of gap that turns a tight placement into an OOM.
Sampling defaults are not the usual ones
The card asks for temperature 1.0 and top_p 1.0, and for reasoning modes a context window of at least 384K. Carrying over a config tuned for a different model is how you get quality regressions that look like the model is worse than advertised.
Serving it
the playbookThe published vLLM recipe targets GB200, H200, B200 and B300 at TP=4 with DP=4, fp8 KV cache, block size 256, and gpu_memory_utilization at 0.92. A single-GPU DGX Station config exists at TP=1, PP=1 with 3-token MTP speculation and FULL_AND_PIECEWISE compilation.
On a 4,000-GPU HGX B200 fleet
500 nodes, eight B200s each, 180 GB per GPU
Weights are 150 GB (FP4 experts + FP8 attention, derived). Usable memory is 180 GB x 0.92 per GPU,
the gpu_memory_utilization from the published recipe. A placement only counts
here if the cache gets at least 30% of what is left, because a config that fits the weights
and nothing else cannot serve a request.
Tensor parallel stays inside the NVSwitch, so no TP traffic ever crosses a NIC.
- Two GPUs is enough, four is the recipe. The derived footprint fits at TP=2 with 55% of memory left for cache, which would give four replicas per node. The published recipe uses TP=4 with DP=4, and until I have measured otherwise I would follow the recipe rather than my own arithmetic.
- It needs its own parsers. Requires --tokenizer-mode deepseek_v4 with the matching --reasoning-parser and --tool-call-parser. Point a V3 config at it and the failures look like a model problem when they are a config problem, which is a bad hour to spend.
- Verified quality. 0.9439 exact match on GSM8K 8-shot, measured on 4x MI355X. Not a Blackwell number, but a real one on a real config.
The one I would put in production
my readThirteen billion active parameters at a million tokens of context is the combination that makes agent workloads stop being a budget conversation. There is a published recipe somebody already tuned, which I value more than another point of benchmark score, and it runs on a single GPU if you need it to. My one caution is the derived weight figure below: I computed it from the published format rather than a byte count, so I would verify the real footprint against the checkpoint before committing a placement to it.
The numbers
measured elsewhere, not here- Total
- 284B
- Active
- 13B
- Output
- 135.2 tok/s
- Output price
- $1.32 / 1M
Throughput from Artificial Analysis on the 0731 reasoning, max-effort endpoint. GSM8K figure from the vLLM recipe on MI355X.
Sources: vLLM Recipes · Hugging Face: DeepSeek-V4-Flash · Artificial Analysis