DeepSeek V4 Flash 0731 (DeepInfra)

The Agent Report describes the July 31, 2026 release of DeepSeek V4-Flash-0731 as a retrained checkpoint of the April preview rather than a new architecture, keeping the same 284-billion-parameter Mixture-of-Experts design with 13 billion active parameters per token. The entire delta is post-training: an expanded reinforcement-learning phase targeting instruction-following, tool-call accuracy under extended context and multi-step chain-of-thought reasoning. According to the piece, the retrained model beats DeepSeek's own V4-Pro-Preview on nine published agent and coding benchmarks, ships with MIT-licensed weights on Hugging Face, and introduces native OpenAI Responses API support for Codex CLI users at $0.14 per million input tokens. The analysis cautions that the headline benchmark numbers carry reproducibility caveats.

DeepSeek published DeepSeek-V4-Flash-0731 to Hugging Face on July 31, 2026 under an MIT license with no access gates, so the weights can be downloaded or deployed without approval. The model card describes substantially enhanced agentic capabilities over the preview version and notes a built-in speculative decoding module. The report records the 48-file safetensors archive at roughly 167 gigabytes, with the model card referencing a 4xGB300 node as one supported hardware configuration. All benchmark figures in the card are vendor-reported across public and internal test sets, with no confirmed independent third-party reproduction, and parameter counts are listed inconsistently at 304B by weight metadata.

Artificial Analysis independently measured DeepSeek V4 Flash 0731 at 50 on its Intelligence Index, a 10-point jump over the April 2026 DeepSeek V4 Flash and one point behind GPT-5.6 Luna (max, 51). The release lands in line with Gemini 3.6 Flash (50) but remains seven points behind the open-weights frontier set by Kimi K3 (max, 57). Agentic gains drive the result: GDPval-AA v2 Elo rose to 1559 from 1189, and Terminal-Bench 2.1 climbed 17 points to 79 percent, a figure below DeepSeek's own reported 82.7. The AA-Omniscience Index improved to -16 purely through a lower hallucination rate of 84 percent, while accuracy stayed flat.

An unaffiliated launch-day audit of DeepSeek V4 Flash 0731 concludes that nearly every published score is vendor-reported, with DeepSeek's own table showing Terminal Bench 2.1 at 82.7, NL2Repo 54.2, Cybergym 76.7 and DeepSWE 54.4 measured with the not-yet-released DeepSeek Harness in minimal mode at maximum reasoning effort. The one same-day independent signal came from Artificial Analysis, which reran its own nine-benchmark harness and measured an Intelligence Index of 50 versus 40 for the prior Flash, with GDPval-AA v2 Elo rising from 1189 to 1559. The audit notes DeepSeek was absent from the Terminal-Bench official leaderboard on July 31, 2026 and that two benchmark rows use internal datasets outsiders cannot rerun.

DeepSeek published DeepSeek-V4-Flash-0731 as the official release of DeepSeek-V4-Flash, superseding the preview with substantially enhanced agentic capabilities. The model keeps the same structure as V4-Flash-DSpark, including an attached speculative decoding module, and DeepSeek reports it beating V4-Pro (Preview) despite far fewer activated parameters. The vendor benchmark table lists Terminal Bench 2.1 at 82.7, NL2Repo 54.2, Cybergym 76.7, DeepSWE 54.4, Toolathlon-Verified 70.3 and Agents' Last Exam 25.2. Developer details include three reasoning-effort levels (low, high, max), no Jinja chat template with dedicated OpenAI-compatible encoding scripts, and vLLM DSpark speculative decoding via a --speculative-config flag with method dspark.