Leadership1 publisher2 min readPublished
DeepSeek's V4.1-Flash reads a million-token prompt on 8B active parameters
The model card for this 552B-parameter Mixture-of-Experts release puts the global KV cache at 890 bytes per token, about a quarter of the previous Flash generation, and every figure in it is DeepSeek's own.
The Board Room · Leadership desk
What happened
- DeepSeek's model card describes V4.1-Flash as a multimodal Mixture-of-Experts model with a 552B-parameter backbone and support for contexts of up to one million tokens.
- The card says the model activates only 8B parameters per token during prefill and 16B during decode, which DeepSeek credits with substantially improving cost efficiency for input-heavy agentic workloads.
- Compressed Sparse Attention 2 plus FP4 caching of the main KV bring the global cache to 890 bytes per token, which the card puts at roughly a quarter of DeepSeek-V4-Flash.
- Pre-training ran on 45 trillion multimodal tokens, with sparse attention trained at a 64K sequence length before the context was extended to one million tokens at the 34 trillion mark.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- capability Holding a full million-token conversation in cache comes to under a gigabyte of memory, so how much context an agent keeps becomes a serving-configuration choice for whoever runs the weights.
- constraint A buyer weighing this against a closed API is putting a published architecture spec next to a metered price, because the cache ratios in the card measure DeepSeek's own generations.
- decision Budgeting a task means setting reasoning_effort somewhere between 1 and 100 before anything can be priced, since the published quality figures all come from the setting the card says costs the most.
- precedent Once a per-token cache budget is a headline spec, procurement teams have a number they can ask every other vendor to match.
Prefill is the phase that reads the repository, the tool output and the transcript so far. Sizing it is where these numbers land. A filled one-million-token window works out to roughly 890 MB of global KV cache at the card's per-token figure [1]. At the four-fold ratio the card claims against DeepSeek-V4-Flash, the same window on the previous generation would have needed about 3.6 GB [3]. The 8B parameters active during prefill are about 1.4 percent of the 552B backbone [2].
Two ratios in the card describe two different caches. The 890 bytes and the four-fold figure cover the global KV cache, which DeepSeek attributes to CSA2's three static attention modes together with FP4 caching of main KV in E2M1 format, one E4M3 scale per 16 channels [6][5]. The one-eighth figure covers the persistent cache: SWA Bounded Replay rebuilds missing sliding-window states by replaying only the most recent n_win tokens, which the card says removes the need to persist SWA KV to SSD [4].
The card is explicit about where the prefill saving comes from. Under the causal encoder-decoder layout, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states, across 20 encoder layers and 20 decoder layers [2]. Each MoE layer carries one shared expert and 384 routed experts, and routes six of them per token [7].
All base models were evaluated in DeepSeek's internal framework under the same settings, with scores within 0.3 of each other treated as equivalent [13]. The cache chart runs across generations of DeepSeek models, so those reductions measure DeepSeek against itself [14].
reasoning_effort runs from 1 to 100 and, in the card's own description, trades inference cost for accuracy [11]. Quality is reported at the setting that costs the most, while the 8B prefill figure holds at any setting [12][3]. The agentic runs used a 1M-token context limit and max_steps=500 per agent, with N=8 samples per task on DeepSWE v1.1 and N=3 on Terminal-Bench 2.1 [15].
The context extension arrived late in training. That leaves 11 trillion of the 45 trillion pre-training tokens seen at the full million-token window [4]. Pricing is what would settle whether open weights now set the cost floor for input-heavy agentic work, and this model card, published in the deepseek-ai namespace on Hugging Face, does not state prices or license terms [16][17].
What to watch
- An independent reproduction of the 890-byte-per-token cache figure outside DeepSeek's internal evaluation framework.
- Serving prices, from DeepSeek or a third-party host, that would let the prefill claim be set against frontier API bills.
- Benchmark results at a low reasoning_effort setting, which would show what the lower setting costs in accuracy.