DeepSeek just shipped DeepSeek V4.1-Flash. The headline items are a 1 million-token context window and, if DeepSeek’s numbers hold up, less overhead when you’re running very long contexts.
DeepSeek says the model uses 890 bytes of global KV cache per token. That’s about 4x smaller than DeepSeek V4-Flash and 437x smaller than DeepSeek V1. The model sits on a 552 billion-parameter backbone, plus 196 billion Engram parameters, with 8 billion activated per token during prefill and 16 billion during decode. DeepSeek also says the 40-layer encoder-decoder split cuts prefill compute by nearly half for the kinds of workloads people actually throw at this stuff: long documents, codebases, and agent workflows.
Most of what’s new comes from Compressed Sparse Attention 2. DeepSeek says it adds Full, Reindex, and Reuse modes, along with KV reuse, FP4 E2M1 main KV, 128-token replay, decoder KV kept in host DRAM rather than SSD, and a 72-hour global KV lifetime. If you’re trying to serve million-token contexts, that’s probably the part you’d want to look at first. Early testers, though, reported uneven multimodal behavior and weaker physics results, even as throughput reached 427 tokens per second, again on DeepSeek’s figures. DeepSeek also lowered DeepSeek V4-Flash pricing while lining up DeepSeek V4.1-Flash against DeepSeek V4 Pro.
You can grab the MIT-licensed weights on Hugging Face for vLLM, SGLang, and Transformers, or use the public API with low, high, and max reasoning tiers.