deepseek-v4-flash-0731 (CanopyWave)

DeepSeek published V4-Flash-0731 on Hugging Face and moved the deepseek-v4-flash API into public beta on July 31, 2026, replacing the April preview without changing the model ID. The 284B-parameter mixture-of-experts architecture is identical to the preview, so the reported performance jump comes entirely from re-post-training, and the checkpoint ships with the DSpark speculative decoding module built in. On DeepSeek's own model-card benchmarks, the 0731 build edges ahead of V4-Pro Preview on every listed agentic task, with Terminal Bench 2.1 at 82.7 versus 61.8 for the preview, DeepSWE at 54.4 versus 7.3, and Toolathlon-Verified at 70.3. The page notes code-agent runs used the unreleased DeepSeek Harness minimal mode and that two benchmark sets are internal, so independent scores may differ.

On July 31, 2026 DeepSeek moved the official V4-Flash API into public beta, with the stable name deepseek-v4-flash now pointing to DeepSeek-V4-Flash-0731. The transformer was unchanged: 43 layers, 284B total parameters, 13B active per token, and a million-token context, while the release added DSpark speculative decoding, reasoning-effort controls, native Responses API support, and Codex adaptation alongside MIT-licensed weights. DeepSWE climbed from 7.3 on the preview to 54.4 on the 0731 checkpoint, though DeepSeek ran that benchmark at maximum reasoning effort using an unreleased "DeepSeek Harness minimal mode" with two of nine launch benchmarks internal. The accompanying DeepSeek-V4 technical report describes million-token MoE models built on hybrid CSA/HCA attention, mHC residual connections, Muon optimization, FP4-aware deployment, and specialist-to-on-policy-distillation post-training.

Artificial Analysis reports that DeepSeek V4 Flash 0731 scores 50 on its Intelligence Index, a 10-point jump over the April 2026 DeepSeek V4 Flash (40) and 6 points ahead of DeepSeek V4 Pro, putting it one point behind GPT-5.6 Luna (max, 51) and level with Gemini 3.6 Flash. The model remains 7 points behind the open-weights leader Kimi K3 (max, 57), and lands on the Intelligence vs Cost per Task Pareto frontier. Agentic gains dominate the evaluation: GDPval-AA v2 Elo rises to 1559 from 1189 for the previous V4 Flash, Terminal-Bench 2.1 climbs 17 points to 79%, and tau3-Bench Banking gains 8 points to 31%. The AA-Omniscience Index improves to -16 from fewer hallucinations rather than higher accuracy, with hallucination rate dropping 12 points to 84%, while the 1M context and 284B/13B-active sizing stay unchanged.

DeepSeek's deepseek-v4-flash-0731 is documented as the official release of DeepSeek-V4-Flash, superseding an earlier preview version with substantially enhanced agentic capabilities. The version-specific documentation lists a 1,048,576-token context window, function calling support, configurable reasoning levels from low to max, batch support, and zero data retention for stored data. Cloudflare's Workers AI documentation exposes the release under the model ID @cf/deepseek-ai/deepseek-v4-flash-0731 as a Cloudflare-hosted deployment, and the page marks the listing as planned for deprecation on 11/20/2026. The page also shows a Workers AI LLM Playground that needs no authentication, so developers can test the model's prompt and message input formats before integrating them.

DeepSeek V4 Flash 0731 is described as an API-only rollout in public beta: the July 31, 2026 update replaces the April preview behind the existing deepseek-v4-flash model ID, so no model-name migration is needed, and DeepSeek V4 Pro, the DeepSeek app, and web chat were not upgraded. Architecture is unchanged, with total parameters, active parameters, and the 1M context window remaining the same. The update adds native Responses API support and adapts the model for Codex, with the release described as a post-training upgrade rather than a new architecture. Flash remains a 284B-parameter mixture-of-experts model with 13B active parameters, and agent performance is the headline of the change.