DeepSeek V4 Flash 0731 (Cloudflare)

DeepSeek's own model card says DeepSeek-V4-Flash-0731 keeps the same architecture and size as the preview build and was only re-post-trained, retaining the 284B-parameter Mixture-of-Experts structure with roughly 13B parameters active per token and a 1M context window. The naming is a date stamp for July 31, 2026, marking a build rather than a new generation, and the weights are published openly on Hugging Face under an MIT license. Agent scores moved sharply: Terminal Bench 2.1 went 61.8 to 82.7 and DeepSWE 7.3 to 54.4, and the model reportedly beats V4-Pro Preview on all nine published agent benchmarks while activating roughly a quarter of Pro's parameters per token. The article stresses that every comparison number comes from DeepSeek's own unreleased harness, so readers should benchmark independently before migrating.

The July 31, 2026 DeepSeek changelog moved V4-Flash out of Preview into official release as DeepSeek-V4-Flash-0731, keeping the same 284B-total and 13B-active MoE architecture and re-post-training only. The update was scoped to the V4-Flash API, with the app, website, and V4-Pro API left untouched, and the changelog markets the smaller model's agent scores as far exceeding the bigger one on DeepSeek's own harness. Operationally, the announcement lists a 2,500 concurrent request ceiling against 500 on V4-Pro, a roughly five-fold headroom increase, while the announced peak and off-peak pricing was not live at publication and the 0731 weights had not yet appeared on Hugging Face. Terminal-Bench 2.1 at 82.7 comes from DeepSeek's own harness and is labeled unverified, whereas the Artificial Analysis Intelligence Index score of 50 is independent.

Artificial Analysis reports DeepSeek V4 Flash 0731 scores 50 on its Intelligence Index, a 10-point gain over the April 2026 DeepSeek V4 Flash and six points ahead of DeepSeek V4 Pro. It lands on the Pareto frontier for intelligence versus cost per task, sitting one point behind GPT-5.6 Luna (max, 51) and in line with Gemini 3.6 Flash (50), while remaining seven points behind the open-weights frontier set by Kimi K3 (max, 57). The evaluation highlights agentic gains: an Elo rating of 1559 on GDPval-AA v2, up from 1189 for the previous V4 Flash, plus a 17-point rise to 79% on Terminal-Bench 2.1 and 31% on τ³-Bench Banking. The 1M token context window, 284B total and 13B active parameters, and a 98% cache-hit discount on DeepSeek's first-party API are also unchanged, and full weights were expected in the coming weeks.

Cloudflare's Workers AI documentation lists the model ID @cf/deepseek-ai/deepseek-v4-flash-0731, describing DeepSeek-V4-Flash-0731 as the official release that supersedes the preview version with substantially enhanced agentic capabilities. The hosted deployment supports function calling, batch requests, and zero data retention, with reasoning levels set to low, high as default, and max, plus a context window of 1,048,576 tokens. The Cloudflare entry is marked for planned deprecation on 11/20/2026, so builders depending on that hosted path should plan for migration. Cloudflare lists host-side unit pricing of $0.44 per million input tokens, $1.32 per million output tokens, and $0.014 per million cached input tokens, and the page documents the request parameters and a no-setup Workers AI LLM Playground for quick testing.

DeepSeek released DeepSeek-V4-Flash-0731 as the official successor to the DeepSeek-V4-Flash preview, adding substantially stronger agentic capability and the same architecture as DeepSeek-V4-Flash-DSpark, including an attached speculative decoding module. The reasoning effort parameter now offers three levels, low, high, and max, controlling how much deliberation the model applies before answering. The model card reports that DeepSeek-V4-Flash-0731 beats DeepSeek-V4-Pro (Preview) on listed benchmarks despite a far smaller activated parameter count, scoring 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, and 54.4 on DeepSWE, with further results against GLM-5.2 and Opus-4.8. The release omits a Jinja chat template and instead supplies encoding scripts for OpenAI-compatible message formatting, and vLLM users enable speculative decoding with a dspark speculative-config flag.

DeepSeek pushed the official deepseek-v4-flash checkpoint into public beta on July 31, 2026, keeping the exact architecture and size of the April preview and rerunning only post-training. The result, by DeepSeek's own published numbers, clears the larger V4-Pro-Preview across all nine agent benchmarks, with the largest swing on DeepSWE going from 7.3 to 54.4, while pricing held flat at $0.14 input and $0.28 output per million tokens with a 98% cache-hit discount. The analysis flags important reproducibility limits: independent Artificial Analysis scoring puts the model at 50 on the Intelligence Index, third of 101 and one point behind GLM-5.2, while flagging a 3.7-point gap between DeepSeek's self-reported Terminal Bench score of 82.7 and its own measurement of 79. The headline agent scores used an unreleased harness, and MIT-licensed weights shipped on Hugging Face with community GGUF quants running on 110 to 168 GB of RAM.