DeepSeek releases V4 Flash 0731 as MIT-licensed open weights

• DeepSeek shipped V4 Flash 0731 as the official Flash release with MIT-licensed open weights, a technical report, API access, and the same model structure as DSpark, according to kimmonismus's model-card post.
• The jump is concentrated in agent and coding tasks: Terminal Bench 2.1 moved from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4 in kimmonismus's benchmark table.
• Artificial Analysis put V4 Flash 0731 on its intelligence-cost Pareto frontier, with a 50 Intelligence Index, 284B total parameters, 13B active parameters, and a 98% cache-hit discount in Artificial Analysis's summary.
• The serving story is unusually concrete: vLLM's note names 256 routed experts, six active per token, and a one-flag DSpark speculative decoding path, while Unsloth's guide says local GGUF runs start around 110GB RAM for 3-bit and 168GB RAM for lossless 4-bit.
• Day-one access spread fast: Cline made it free in its coding agent in Cline's rollout, and Ollama's cloud launch exposed it through ollama run deepseek-v4-flash:0731-cloud.

The Hugging Face model card says 0731 supersedes the preview, keeps DSpark attached, and replaces the usual Jinja chat template with explicit encoding and parsing scripts. DeepSeek's API docs say the existing deepseek-v4-flash slug now resolves to 0731 through the same call path. Unsloth's GGUF page puts the lossless Q8 build at 162GB, only 7GB larger than Q4. On Hacker News, the immediate nerd snipe was the unreleased DeepSeek Harness used for code-agent benchmark tasks.

DeepSeek V4 Flash 0731 started as an API update, then the weights followed. The release is unusually clean for an open-weight model: no subscription launch theater, just a model card, a license badge, and runnable artifacts.

• API: the existing deepseek-v4-flash model name now points at 0731, and Cline's changelog link captured the initial API-only state before the weights landed.
• Open weights: Artificial Analysis's open-weights note says the weights are MIT licensed and released in mixed FP4/FP8 precision at roughly 167GB.
• Scope: the changelog screenshot in Zhihu Frontier's post says V4-Pro API and the app or web models were not changed in this update.
• Context: Artificial Analysis's summary lists a 1M-token context window, unchanged from the earlier Flash.
• Pricing: Artificial Analysis's summary lists $0.14 per 1M input tokens, $0.28 per 1M output tokens, and $0.0028 per 1M cache-hit tokens.

First-party

• Terminal Bench 2.1: 61.8 → 82.7, +20.9 points, in kimmonismus's benchmark table.
• DeepSWE: 7.3 → 54.4, +47.1 points, in kimmonismus's benchmark table.
• Cybergym: 38.7 → 76.7, +38.0 points, in kimmonismus's benchmark table.
• DSBench-FullStack: 37.0 → 68.7, +31.7 points, in kimmonismus's benchmark table.

Third-party evaluators

• Artificial Analysis Intelligence Index: 40 → 50, +10 points, per Artificial Analysis's summary.
• GDPval-AA v2: 1189 → 1559 Elo, +370 Elo, per Artificial Analysis's GDPval chart.
• Terminal-Bench 2.1 in the Artificial Analysis suite: 62% → 79%, +17 points, per Artificial Analysis's summary.
• Frontend Code Arena: 1432 → 1586, +154 points, according to Arena's leaderboard post.

Customer-reported

• Data-extraction recall: 71% → 79%, +8 points, on a long-running production job described in rishdotblog's hands-on report.

No cited eval in the evidence shows a predecessor regression. The caveats are absolute weaknesses, benchmark opacity, and missing modalities.

• AA-Omniscience accuracy stayed 37% → 37%, +0 points, while the hallucination rate fell 96% → 84%, -12 points, according to Artificial Analysis's summary.
• CritPt improved to 17%, but teortaxesTex's CritPt note placed it below GLM-5.2, Luna, and K3 on that physics-reasoning slice.
• The release is text-only, and Artificial Analysis's summary lists no image input or multimodal support.
• The public code-agent numbers used an upcoming DeepSeek Harness in minimal mode, and mattlam_'s eval comment argued that closed eval harnesses are hard to trust until others can inspect them.
• The model still acts blind in some visual-agent runs, with teortaxesTex's pixel-forensics post describing repeated attempts to reconstruct vision through programmatic image analysis.

The boring line is the important one: the changelog captured in Zhihu Frontier's post says the model structure and size stayed the same, and the release was re-post-trained.

• MoE shape: 284B total parameters and 13B active at inference time, per Artificial Analysis's summary.
• Routing: 256 routed experts with six active per token, according to vLLM's note.
• Speculative decoding: DSpark ships inside the weights, so vLLM can enable it with --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}', per vLLM's note.
• Reasoning effort: the model card described three levels, low, high, and max, and Simon Willison's post found the pelican test improved when reasoning_effort moved to high.
• Serving contract: the Hugging Face model card says there is no Jinja-format chat template; TheZachMueller's serving-contract note says OpenAI tool messages need conversion into DeepSeek's DSML format and parsed back into reasoning, content, and tool calls.
• Kernels: maharshii's TileLang note spotted DeepSeek using TileLang for an FP8 activation and FP4 weight GEMM kernel.

Hands-on reports were more useful than the launch adjectives. They exposed capacity limits, reasoning-mode sensitivity, long-context behavior, and one very weird house style.

• A production extraction job saw recall rise from 71% to 79%, while rishdotblog's capacity reply reported 503 "service too busy" errors during the rollout.
• rishdotblog's usage screenshot showed 1.07B DeepSeek tokens over five days for SEC filings and earnings-transcript cleanup, at $198.98 in API cost.
• Default reasoning gave Simon Willison's pelican test a broken bicycle scene; the high-effort rerun produced a cleaner SVG, with Willison's cost reply putting the two calls at fractions of a cent.
• Long context looked usable deep into a run: teortaxesTex's context screenshot reported 475K tokens in with no clear fatigue.
• Open-book QA still had blind spots, but teortaxesTex's QA follow-up said 0731 could recover after feedback where the preview failed.
• The reasoning trace has a clipped "Grug Speech" style in English, visible in teortaxesTex's comparison screenshot.

The rollout moved from model card to tools within hours.

• Cline made V4 Flash 0731 free and exposed it through Settings → Cline provider → deepseek-v4-flash, per Cline's rollout.
• Ollama added a cloud route with ollama run deepseek-v4-flash:0731-cloud, and Ollama's post also showed a Claude Code launch command.
• Arena added it to Agent Mode and Battle Mode, with Arena's post describing long-horizon tasks with web search, filesystem, and terminal tools.
• vLLM published a one-flag DSpark serving path in vLLM's note.
• Unsloth shipped local GGUF runs, then Unsloth's smaller-quant update added 1-bit through 4-bit variants.
• Venice made the model available in a private chat surface, according to Venice's post.
• Hermes Agent initially announced Nous Portal and OpenRouter support, but Teknium's follow-up said Nous Portal blocked providers that train on user data until alternate providers appeared; Teknium's resolution note said the route was fixed once more providers came online.