DeepSeek quietly slipped V4-Flash-0731 into public beta on July 31, 2026. No press event, no livestream, just an API changelog entry. But within hours, the AI developer community caught on to what had happened: a model with one-sixth the total parameters of its flagship sibling had just beaten it across every single agent and coding benchmark DeepSeek tested.
The update keeps the exact same 284B total parameter, 13B active parameter MoE architecture from the April preview. No new layers, no bigger context window, no expanded model size. The only thing that changed was post-training. And yet DeepSWE jumped from 7.3 to 54.4, Terminal Bench 2.1 hit 82.7, and the Artificial Analysis Intelligence Index landed at 50, one point behind GPT-5.6 Luna.
If you track AI model releases, that pattern should catch your attention. DeepSeek proved that the gap between "good enough" and "frontier-tier" can close without adding parameters.
DeepSeek V4 Flash 0731: Version Highlights at a Glance
The table below summarizes the key DeepSeek V4 Flash 0731 update details for quick reference.
| Category | Details |
|---|---|
| Model name | DeepSeek-V4-Flash-0731 |
| Release date | July 31, 2026 |
| Status | Public beta (API only) |
| Architecture | MoE, 284B total / 13B active params |
| Context window | 1M tokens |
| Max output | 384K tokens |
| License | MIT (open weight on HuggingFace) |
| Key change | Post-training only, no architecture change |
| New API support | Native Responses API, Codex adaptation |
| Pricing | $0.14/M input, $0.28/M output |
Same Skeleton, New Brain: How Post-Training Drove the DeepSeek Update
Here is what makes this DeepSeek API update unusual. Most model upgrades come with bigger parameter counts, expanded context, or new architectural tricks. V4-Flash-0731 did none of that.
The model structure, parameter count, and 1M token context window are identical to the V4-Flash Preview released in April. DeepSeek confirmed in their changelog that the only change was a full re-run of post-training. That means new supervised fine-tuning data, refined reinforcement learning strategies, and what DeepSeek calls their self-developed testing framework for agent task optimization.
The practical takeaway? DeepSeek's post-training pipeline has gotten good enough that the same model can go from "barely functional at agent tasks" to "competitive with frontier models" without anyone touching the underlying weights or architecture. DeepSWE went from 7.3 to 54.4. That's not a marginal improvement. It's a model that couldn't do software engineering tasks three months ago now scoring above 54 on a benchmark that requires reading code repositories, locating problems, modifying multiple files, running tests, and fixing errors based on output.
This matters for developers because it suggests the bottleneck in AI coding is moving. Pre-training scale used to be the ceiling. Post-training quality is becoming the lever.
9 Benchmarks Where DeepSeek V4 Flash Flipped the Script
DeepSeek published full benchmark numbers in their official model card on HuggingFace, and they did something rare in the AI industry: they put their Flash model head-to-head against their own V4 Pro Preview in a direct comparison table. The smaller model won every single category.
The table below breaks down the DeepSeek V4 benchmark results across all 9 agent and coding tests.
| Benchmark | V4-Flash Preview | V4-Flash 0731 | V4-Pro Preview | Claude Opus 4.8 |
|---|---|---|---|---|
| Terminal Bench 2.1 | 61.8 | 82.7 | 72.1 | 85.0 |
| NL2Repo | 39.4 | 54.2 | N/A | N/A |
| Cybergym | 38.7 | 76.7 | N/A | N/A |
| DeepSWE | 7.3 | 54.4 | N/A | N/A |
| Toolathlon Verified | N/A | 70.3 | N/A | N/A |
| Agent Last Exam | N/A | 25.2 | 15.8 | 25.7 |
| Automation Bench (Public) | N/A | 25.1 | N/A | N/A |
| DSBench-FullStack | 37.0 | 68.7 | N/A | 71.6 |
| DSBench-Hard | 25.8 | 59.6 | N/A | 71.7 |
A few things jump out from these numbers.
The DeepSWE jump from 7.3 to 54.4 is the most dramatic single improvement. That's a 645% increase on a benchmark that measures long-horizon software engineering: understanding a codebase, finding the right files, making multi-file changes, and verifying fixes through test execution. The preview version was essentially non-functional on these tasks. The formal version is now within striking distance of dedicated coding models.
On Terminal Bench 2.1, V4-Flash 0731 scored 82.7, beating GLM-5.2 (81.0) and trailing Claude Opus 4.8 (85.0) by just 2.3 points. Terminal Bench tests whether a model can operate a sandboxed terminal: running commands, reading output, and adjusting actions based on results. That's the difference between a model that writes code and a model that engineers software.
The Agent Last Exam score of 25.2 deserves attention too. Claude Opus 4.8 scored 25.7. Half a point. DeepSeek V4 Flash came within half a point of one of the most capable frontier models on a benchmark designed to test whether AI agents can complete real-world tasks rather than answer trivia.
DSBench-FullStack at 68.7 is close to Opus 4.8's 71.6. DSBench-Hard at 59.6 still has a 12-point gap to Opus, so there is room to improve on the hardest coding agent tasks. But the trend line is clear.
One caveat worth noting: these are official scores from DeepSeek's own evaluation environment using their testing framework in minimal mode with max reasoning effort. Independent testing from Artificial Analysis showed a slightly lower Terminal Bench score of 79%, which is still strong but not identical to the official 82.7. The gap between official and third-party scores is normal for new model releases, but it means you should treat these numbers as upper-bound estimates until more independent benchmarks come in.
DeepSeek V4 Pro vs V4 Flash: Why the Smaller Model Won
The comparison between DeepSeek V4 Pro and V4 Flash is where this update gets genuinely surprising. V4 Pro Preview has 1.6 trillion total parameters and 490 billion active parameters per token. V4 Flash has 284 billion total and 13 billion active. Pro is roughly 5.6x bigger in total parameters and 3.8x bigger in active parameters.
And yet Flash 0731 beat Pro Preview on all 9 benchmarks DeepSeek tested.
How? The answer is entirely in the post-training. V4 Pro Preview was released in April with strong general reasoning but relatively untuned agent capabilities. V4 Flash Preview shipped at the same time with even weaker agent performance. Over three months, DeepSeek refined the Flash model's post-training to target exactly the capabilities that matter for coding agents: tool calling, multi-step execution, terminal operations, and code repository navigation.
The implication for the broader AI field is clear. If post-training can close a 5x parameter gap, the arms race for bigger models may be shifting toward a race for better training data and strategies. DeepSeek V4 Pro's formal release is expected in August 2026, and if it gets the same post-training treatment, the results could shake up the leaderboard.
Codex Integration: What the DeepSeek API Update Means for Developers
Alongside the benchmark improvements, DeepSeek V4 Flash 0731 added native support for OpenAI's Responses API format. This is the interface that Codex uses to communicate with models, and it matters more than it sounds.
Before this update, developers who wanted to use DeepSeek with Codex needed intermediary tools like CC Switch or Moon Bridge to translate between DeepSeek's native Chat Completions API and Codex's Responses API format. Now V4 Flash speaks Responses API directly. You can configure DeepSeek as a model provider in Codex CLI, the ChatGPT desktop app, and the VS Code Codex extension with a single setup command.
DeepSeek provided automatic setup scripts for this. On macOS and Linux:
bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.sh)
On Windows PowerShell:
irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex
The Responses API currently supports only deepseek-v4-flash. V4 Pro is expected to add Responses API support in early August.
For developers building coding agents, this matters because it removes a layer of friction. You can drop DeepSeek directly into a Codex workflow without maintaining a proxy or bridge service. The model name stays deepseek-v4-flash and the base URL doesn't change, so existing integrations keep working with zero code modification.
Third-Party Verification: Artificial Analysis and Arena.ai Confirm DeepSeek V4 Benchmark Results
Two independent benchmark platforms confirmed the V4 Flash 0731 update results within 24 hours of release.
Artificial Analysis gave V4 Flash 0731 an Intelligence Index score of 50 out of 100. For context, the previous Flash Preview scored 40, and V4 Pro Preview scored 44. The median score among comparable models is 17. GPT-5.6 Luna sits at 51, making V4 Flash just 1 point behind OpenAI's current flagship on this index.
The cost-per-task numbers are where DeepSeek continues to dominate. Artificial Analysis calculated V4 Flash's average cost per Intelligence Index task at $0.03. For comparison, Kimi K3 costs $0.86 per task, GPT-5.6 Sol costs $1.86, and Claude Fable 5 costs $3.15. Even after OpenAI cut GPT-5.6 Luna's price by 80%, DeepSeek's per-task cost remains about 60% lower, thanks to a 98% cache hit discount on DeepSeek's own API.
Arena.ai's Frontend Code Arena, which evaluates models on web development and automated coding through blind human preference testing, ranked V4-Flash-High at 1586 Elo. That's a 154-point jump from the Flash Preview and a 121-point jump from V4 Pro Preview. It ranks 3rd in the open-weight category and 7th overall.
These third-party results validate the official numbers to a degree. The Artificial Analysis Intelligence Index of 50 aligns with DeepSeek's own claims about approaching frontier-tier performance. The Arena.ai ranking, based on human evaluations rather than automated benchmarks, adds a layer of credibility that official benchmarks alone can't provide.
What This DeepSeek New Version Means for App Users
Here is an important detail: the V4 Flash 0731 update only applies to the API. The DeepSeek app and web interface still run on the previous model version. If you use DeepSeek through its Android app or website, you won't notice any change yet.
DeepSeek has stated that V4 Pro's formal release is coming in August, and the app and web models will be updated around that time. The Responses API support is also limited to V4 Flash for now, with V4 Pro expected to gain support shortly.
For developers using the API, the transition is automatic. The model name deepseek-v4-flash now points to the 0731 version with no code changes needed. Pricing stays the same: $0.14 per million input tokens, $0.28 per million output tokens, with cache hit discounts bringing input costs down to roughly $0.003 per million tokens.
The open-weight release on HuggingFace under MIT license means you can also self-host V4 Flash 0731. The model files come in at around 167GB with FP4/FP8 mixed precision. That's large for the "lightweight" model in DeepSeek's lineup, but the 13B active parameter count means inference is far cheaper than running a dense model of comparable capability.
Conclusion: DeepSeek V4 Flash Update Signals a Post-Training Shift
DeepSeek V4 Flash 0731 proves that post-training can do what parameter scaling used to do: close the gap with frontier models. A 284B model with 13B active parameters beating a 1.6T model with 49B active parameters across 9 benchmarks is a clear signal that the industry's focus is shifting from "bigger is better" to "smarter training wins." For developers, the V4 Flash update brings Codex-native integration, frontier-adjacent coding agent performance, and the same aggressive pricing that made DeepSeek a global API leader. The V4 Pro formal release in August will show whether the flagship model gets the same post-training treatment. If it does, the gap with Claude and GPT could close further than anyone expected.
You can download the DeepSeek app on APKPure to use the AI assistant on your Android device.
DeepSeek is an AI assistant app that offers free access to its latest reasoning model.