DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains

DeepSeek revealed DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API into public beta on July 31, 2026. The mannequin card is express that that is the official launch superseding the preview, and that the structure and measurement are unchanged. The positive aspects come from re-post-training, not a brand new design.

The checkpoint ships with the DSpark speculative decoding module connected, matching the construction of DeepSeek-V4-Flash-DSpark. Hugging Face stories 304B parameters for the repo, which incorporates that draft module on prime of the 284B base.

On the API aspect, deepseek-v4-flash now natively helps the Responses API format and is tailored for Codex. The V4-Pro API and the app and net fashions weren’t up to date.

Is it deployable?

Yes, in two very alternative ways.

Via API, it’s deployable by nearly anybody: DeepSeek’s pricing page lists deepseek-v4-flash at $0.14 per 1M enter tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per 1M output tokens, with a 2,500 concurrency restrict. That is roughly a 3rd of deepseek-v4-pro output pricing ($0.87). Seed-stage startups, indie builders, and inside platform groups can run agent loops at this worth with out a GPU price range.

Via self-hosting, the bar is way increased: The weights are MIT-licensed and ungated, however each skilled stays resident in reminiscence regardless that solely 13B activate per token. DeepSeek’s vLLM instance serves it on a single 4×GB300 node. Unsloth’s dynamic GGUFs put the lossless 8-bit construct at 162 GB and a 3-bit construct at 103 GB, needing roughly 110 GB of mixed RAM plus VRAM. Self-hosting fits mid-size and massive enterprises with a serving cluster, or one well-specced workstation at aggressive quantization.

Architecture

Per the DeepSeek-V4 technical report, V4-Flash is a 284B-parameter MoE with 13B activated per token and a 1M-token context window. Each MoE layer holds 1 shared skilled and 256 routed specialists with an intermediate dimension of 2048, and 6 routed specialists fireplace per token. The first three MoE layers use hash routing. Multi-token prediction depth is 1.

Attention is hybrid, combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Manifold-Constrained Hyper-Connections (mHC) change typical residual connections, with enlargement issue 4 and 20 Sinkhorn-Knopp iterations. Pre-training used greater than 32T tokens and the Muon optimizer. The paper’s headline effectivity determine — 27% of single-token inference FLOPs and 10% of KV cache versus DeepSeek-V3.2 at 1M context — is acknowledged for V4-Pro, not Flash.

<!– EMBED HERE: paste wordpress-embed.html right into a Custom HTML block –>