DeepSeek releases V4.1-Flash, a new open-weight model using a novel causal encoder-decoder architecture with separate 8B prefill / 16B decode active parameters

DeepSeek released V4.1-Flash, replacing V4 Pro as its flagship open-weight model. It uses a new causal encoder-decoder architecture that separates active parameters for prefill and decode (763B total, 8B active on input, 16B active on output), supports 1M-token context and native text plus image input, and ships under an MIT licence. Pricing is $0.30 per 1M input tokens and $1.20 per 1M output tokens, with day-0 hosting from Baseten and a rollout on Ollama's paid tiers. It scored 40 on the Artificial Analysis Intelligence Index and placed first among open-weight models on the Vals Index.

Developers and enterprises building long-running AI agents get a cheaper, MIT-licensed model with a claimed KV cache footprint as low as 1/8 of the previous version, which lowers the cost of maintaining long conversations or agent state. Because it is open-weight with immediate third-party hosting support, teams can test and deploy it today rather than waiting on an announcement. The benchmark score of 40 on the Intelligence Index is middling, so the practical case for switching rests more on cost and context efficiency than on raw capability gains.

The "Return of the Whale" framing and commentator claims that this is effectively a "DeepSeek V5" undersold by naming are promotional spin, not evidence. The article leans on unfalsifiable claims that benchmarks fail to capture the architecture's real advance, which conveniently explains away the fact that the model trails some open competitors on standard measures. No primary source links (paper, docs, benchmark pages) are included despite claiming research-paper and benchmark evidence, so the specific numbers cannot be independently checked from this coverage alone.

1. 01Independent third-party benchmarks (e.g. Artificial Analysis, Vals AI, LMSYS) confirming or revising the Intelligence Index score of 40 and the #1 open-weight ranking over the next 2-4 weeks
2. 02Whether the claimed 1/8 KV cache reduction and prefill/decode split hold up in independent inference benchmarks from hosting providers beyond Baseten and Ollama
3. 03Developer and enterprise adoption signals: GitHub stars, Hugging Face downloads, and whether major inference platforms (Together, Fireworks, Replicate) add support within the next month
4. 04Whether DeepSeek or independent researchers publish the underlying research paper with full architecture details, since none was linked in this coverage