DeepSeek Releases V4 Flash With Agentic Gains
DeepSeek released DeepSeek-V4-Flash-0731 on July 31, superseding the preview model with stronger reported agentic results and native Responses API support. The company says the update changes only the Flash API, while V4 Pro and its app and web models remain unchanged. Artificial Analysis scored Flash-0731 at 50 on its Intelligence Index, 10 points above the previous V4 Flash.
DeepSeek released DeepSeek-V4-Flash-0731 on July 31 as the official version of V4 Flash, superseding the preview release. The public-beta API remains available under the deepseek-v4-flash model name, and DeepSeek says the model was re-post-trained without changing the preview architecture or size.
What changed
DeepSeek's release note reports substantially stronger agent capabilities and native support for the Responses API format. The company says this update applies only to the V4 Flash API: V4 Pro and the models used in DeepSeek's app and website are unchanged. Its Responses API documentation currently supports only deepseek-v4-flash, with V4 Pro support expected in early August 2026.
The official model card says Flash-0731 uses the same model structure as the preview, including a speculative-decoding module. It also exposes low, high, and max reasoning-effort levels. DeepSeek recommends temperature 1.0 and top-p 0.95 for agentic scenarios, while noting that high and max effort can produce very long outputs.
Company-reported benchmark gains
DeepSeek's model card reports 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. Each result is above the V4 Pro preview score in DeepSeek's comparison table.
The evaluation conditions matter. DeepSeek says its public code-agent benchmarks used an unreleased minimal mode of DeepSeek Harness, max reasoning effort, temperature 1.0, and top-p 0.95. It also identifies DSBench-FullStack and DSBench-Hard as internal test sets. Those qualifications make the figures useful as vendor evidence, not substitutes for independent workload testing.
Independent measurements
Artificial Analysis scored Flash-0731 at 50 on its Intelligence Index, up from 40 for the previous V4 Flash. It reports a 1-million-token context window and an unchanged 284-billion-total, 13-billion-active parameter configuration. On GDPval-AA v2, its agentic work benchmark, Flash-0731 rose from 1,189 to 1,559 Elo.
The firm's AA-Omniscience evaluation found a lower hallucination rate without higher accuracy: the rate fell 11 points to 84%, while accuracy remained 37%. That result is specific to the evaluator's test design and should not be generalized to every task.
For production teams, the practical change is an explicit new model version with stronger reported agent performance and Responses API compatibility. Version pinning, regression tests, tool-call validation, latency measurement, and workload-specific quality checks remain the safest basis for adoption.
Key Points
- 1DeepSeek-V4-Flash-0731 supersedes the preview while retaining its architecture and size and adding native Responses API support.
- 2DeepSeek reports large coding-agent benchmark gains, but some evaluation infrastructure and test sets are unreleased or internal.
- 3Artificial Analysis scored the model 10 Intelligence Index points above the previous V4 Flash and reported stronger agentic-work results.
Scoring Rationale
This is a significant frontier-adjacent model update with reported gains in coding-agent performance, long context, and cost efficiency. It is especially relevant to teams operating OpenAI-compatible APIs or evaluating lower-cost models for agent workflows, although performance claims require workload-specific validation.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
