DeepSeek V4 Flash 0731 (NovitaAI)

Artificial Analysis independently scored DeepSeek V4 Flash 0731 at 50 on its Intelligence Index, 10 points above the earlier V4 Flash and one point behind GPT-5.6 Luna at max. Its agentic GDPval-AA v2 Elo rose to 1559 from 1189, while Terminal-Bench 2.1 increased 17 points to 79% and τ³-Bench Banking increased eight points to 31%. The evaluation says the checkpoint retains a 1M-token context window and the prior model’s 284B total and 13B active parameters. Its AA-Omniscience Index improved by seven points to -16 because the hallucination rate fell 12 points to 84%, while overall accuracy remained unchanged, and the model reached the analysis provider’s Intelligence-versus-cost Pareto frontier.

DeepSeek’s official model card names DeepSeek-V4-Flash-0731 as the release that supersedes the V4 Flash preview, with substantially improved agentic capabilities and the same structure as V4-Flash-DSpark, including an attached speculative decoding module. The vendor reports 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. For developers, the checkpoint supports low, high, and max reasoning effort but does not include a Jinja chat template; DeepSeek instead provides encoding scripts for OpenAI-compatible messages and response parsing. Its vLLM guide enables DSpark speculative decoding with a seven-token speculative configuration, while the benchmark table uses DeepSeek’s unreleased Harness and includes internal DSBench datasets.