DeepSeek V4-Flash-0731: A Retrained Build That Outpaces Its Own Flagship
On 31 July 2026, DeepSeek released the official public beta of its V4-Flash API under the build designation DeepSeek-V4-Flash-0731. The release attracted immediate attention not for introducing a new architecture but for what the updated build achieved with the same one: across all nine agent and coding benchmarks that DeepSeek published at launch, V4-Flash-0731 scored higher than DeepSeek-V4-Pro-Preview — the company's own flagship model. A specialised Flash variant, designed for speed and cost efficiency, had overtaken the Pro line on the benchmarks that matter most for agentic and engineering use cases. The practical implication is significant for any team running agentic workflows: the cheaper model now produces better results on the tasks those workflows depend on.
What the Benchmarks Show
The benchmark results published by DeepSeek on 31 July 2026 cover agent interaction, coding, and multi-step task domains. On Terminal-Bench 2.1, which measures performance on complex terminal-based agent tasks, V4-Flash-0731 scored 82.7 — up from 61.8 for the V4-Flash-Preview build and above the 72.1 recorded by V4-Pro-Preview. On DeepSWE, a software engineering benchmark that tests models on real-world codebase tasks, the 0731 build scored 54.4 against 7.3 for the preview build. Cybergym rose to 76.7 from 38.7. NL2Repo, which measures the ability to write code from natural-language descriptions into a repository context, reached 54.2 against 39.4 for the preview. DSBench-FullStack reached 68.7 and DSBench-Hard reached 59.6. These improvements span terminal agents, software engineering, cybersecurity-adjacent tasks, and multi-file coding — not a single synthetic benchmark.
Architecture: Same Model, Upgraded Training
DeepSeek is explicit that V4-Flash-0731 is not a new architecture. It retains the same structure as V4-Flash-Preview: a Mixture-of-Experts design with 284 billion total parameters and 13 billion parameters active per inference token, a one-million-token context window, and the MIT licence that permits self-hosting and commercial deployment without restriction. The upgrade is in post-training. DeepSeek has not published full details of the post-training methodology, but the benchmark improvements across agentic and multi-step task domains suggest that the updated reinforcement learning and fine-tuning runs were specifically targeted at those capabilities rather than at raw language modelling performance.
Pricing and the API Surface
The V4-Flash-0731 API is priced at 0.14 US dollars per one million input tokens and 0.28 US dollars per one million output tokens. DeepSeek's automatic context caching reduces cache-hit input cost to 0.0028 US dollars per one million tokens — a 98 per cent reduction for repeated prompt contexts, with no configuration required and no separate storage charge. Third-party pricing analyses place V4-Flash-0731 at roughly 35 to 100 times cheaper per token than GPT-5.5 or Claude Opus 4.8 at comparable context lengths. The API now natively supports the Responses format used by OpenAI's Codex, meaning applications already built to that interface can route requests to V4-Flash-0731 without changes to the tool-calling or multi-step task logic.
The DeepSWE Score Gap
The jump from preview to 0731 on DeepSWE — from 7.3 to 54.4 — is the most striking result in the published set. DeepSWE tests models on real software engineering tasks drawn from public repositories: the model must understand an existing codebase, plan an implementation, write correct code, and have that code pass tests it did not write. A score of 7.3 indicated the preview build could not reliably complete those tasks end-to-end. A score of 54.4 indicates the 0731 build can complete roughly half of them correctly. That is the difference between a model that assists at the function level and one that can function as a contributor to a real codebase task. The DeepSWE improvement, more than any other benchmark in the published set, explains why this release is being treated as a material capability event rather than a routine incremental update.
What This Means for Indian Engineering Teams
For Indian software and AI product teams, V4-Flash-0731 opens three practical options. First, for teams running agentic coding pipelines — automated code review, test generation, or issue triage — replacing a closed frontier model with V4-Flash-0731 can reduce inference cost by one to two orders of magnitude while improving benchmark performance on the relevant tasks. Second, the MIT licence permits self-hosted deployment on Indian cloud or on-premise infrastructure, which matters for teams handling client code under confidentiality agreements, government data localisation requirements, or DPDPA-governed user data. Third, the Responses API compatibility means the migration path from existing Codex-based workflows is a configuration change rather than a rewrite. For Indian AI product companies competing on deployment cost, the combination of Pro-level agentic performance at Flash-level pricing changes the unit economics of any feature delivered on a per-seat or per-task billing model.
The Bottom Line
On 31 July 2026, DeepSeek released DeepSeek-V4-Flash-0731, a retrained build of its V4-Flash model that outperformed DeepSeek-V4-Pro-Preview on all nine published agent and coding benchmarks. Key scores include 82.7 on Terminal-Bench 2.1, 54.4 on DeepSWE, and 76.7 on Cybergym — improvements of between 15 and 47 percentage points over the preview build. The model retains the same 284 billion parameter Mixture-of-Experts architecture with a one-million-token context window and MIT licence. API pricing is 0.14 US dollars per one million input tokens, with cache hits at 0.0028 per one million tokens. For Indian engineering teams, V4-Flash-0731 is a materially improved and significantly cheaper option for agentic coding and multi-step task workflows than any closed frontier model currently on the market.
Frequently Asked Questions
What is DeepSeek V4-Flash-0731 and when was it released?+
DeepSeek V4-Flash-0731 is the official public beta release of DeepSeek's V4-Flash API, released on 31 July 2026. It is a retrained version of the V4-Flash-Preview model — the architecture is unchanged at 284 billion total parameters with 13 billion active per token in a Mixture-of-Experts design, a one-million-token context window, and an MIT licence. The upgrade is in post-training: updated reinforcement learning and fine-tuning runs produced significant benchmark improvements across agentic and coding domains without changing the model's structural design. The 0731 build scored higher than DeepSeek's own V4-Pro-Preview on all nine published agent and coding benchmarks, including an improvement on DeepSWE from 7.3 to 54.4.
How does DeepSeek V4-Flash-0731 compare to its predecessor on benchmarks?+
DeepSeek V4-Flash-0731 substantially outperforms the V4-Flash-Preview build across all nine published agent and coding benchmarks. On Terminal-Bench 2.1, it scored 82.7 versus 61.8 for the preview. On DeepSWE, the score rose from 7.3 to 54.4 — the largest single improvement. Cybergym rose from 38.7 to 76.7. NL2Repo improved from 39.4 to 54.2. DSBench-FullStack reached 68.7 and DSBench-Hard reached 59.6. On all nine benchmarks, V4-Flash-0731 also outperformed DeepSeek-V4-Pro-Preview, the company's own flagship model, which had recorded 72.1 on Terminal-Bench 2.1. These improvements are attributed to updated reinforcement learning and supervised fine-tuning in the new training run, with DeepSeek explicitly stating that the architecture and parameter count did not change.
What is the API pricing for DeepSeek V4-Flash-0731?+
DeepSeek V4-Flash-0731 is priced at 0.14 US dollars per one million input tokens and 0.28 US dollars per one million output tokens through DeepSeek's official API. Context caching is automatic: cache-hit input tokens cost 0.0028 US dollars per one million tokens — a 98 per cent reduction — with no configuration required and no separate storage or write charge. Third-party pricing analyses place V4-Flash-0731 at roughly 35 to 100 times cheaper per token than comparable frontier models from other providers at similar context lengths. The model is also MIT-licensed, which permits self-hosting on private infrastructure for teams with high-volume agentic workloads.
Can Indian engineering teams self-host DeepSeek V4-Flash-0731?+
Yes. DeepSeek V4-Flash-0731 is released under the MIT licence, which permits commercial deployment and self-hosting on private infrastructure without restriction. The model has 284 billion total parameters with only 13 billion active per token in its Mixture-of-Experts design, which reduces the inference hardware requirement compared to a dense model of similar capability. Indian software and AI product teams can deploy V4-Flash-0731 on Indian cloud infrastructure or on-premise servers, which is relevant for DPDPA compliance, client data residency requirements under government contracts, and confidentiality obligations for client codebases handled under NDA. The model supports a one-million-token context window in self-hosted deployments and the API surface supports the Responses format for Codex-compatible tooling.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.