DeepSeek V4-Flash-0731: Better Agents, Same Architecture, Cheaper API
DeepSeek released V4-Flash-0731 on July 31, 2026. Same 284B architecture, but re-post-training delivers big agentic and coding gains at $0.28/1M output tokens.
DeepSeek published V4-Flash-0731 on Hugging Face and pushed the deepseek-v4-flash API into public beta on July 31, 2026. The 284B-parameter mixture-of-experts model is architecturally identical to the April preview. The performance jump, which sees the model beat V4-Pro Preview on every agentic benchmark DeepSeek published, comes entirely from re-post-training. The API costs $0.28 per million output tokens, roughly a third of V4-Pro, making sustained agent loops affordable even without a GPU budget.
What happened
| Detail | Value |
|---|---|
| Release date | July 31, 2026 |
| Total parameters (with DSpark draft module) | 304B |
| Base model parameters | 284B |
| Activated parameters per token | 13B |
| Context window | 1M tokens |
| API output price | $0.28 per 1M tokens |
| API input price (cache miss) | $0.14 per 1M tokens |
| API input price (cache hit) | $0.0028 per 1M tokens |
| API concurrency limit | 2,500 |
| License | MIT, ungated |
DeepSeek released V4-Flash-0731 as the official successor to the April preview. The model card states clearly that the architecture and size did not change. The checkpoint ships with the DSpark speculative decoding module built in, which is why Hugging Face reports 304B total parameters even though the base model is 284B. Speculative decoding (a technique where a small draft model proposes tokens that the main model then verifies in bulk) is enabled in vLLM with a single flag.
On the API side, deepseek-v4-flash now supports the Responses API format and is adapted for Codex. The V4-Pro API, plus the app and web models, were not updated in this release.
How does V4-Flash-0731 compare to V4-Pro and the preview?
According to DeepSeek’s own model card benchmarks, the 0731 update is a large jump over the April preview and edges ahead of V4-Pro Preview on every agentic task listed:
| Benchmark | V4-Flash-0731 | V4-Flash Preview | V4-Pro Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 |
| NL2Repo | 54.2 | 39.4 | 38.5 |
| Cybergym | 76.7 | 38.7 | 52.7 |
| DeepSWE | 54.4 | 7.3 | 12.8 |
| Toolathlon-Verified | 70.3 | 49.7 | 55.9 |
| Agents’ Last Exam | 25.2 | 15.8 | 16.5 |
| AutomationBench Public | 25.1 | 10.8 | 12.8 |
One important caveat: code agent tasks used the “minimal mode” of DeepSeek Harness, a framework that has not been publicly released. DSBench-FullStack (68.7) and DSBench-Hard (59.6) are internal test sets. Agent benchmark scores are sensitive to how the harness is configured, so numbers from independent runs may differ significantly.
Deployment: API vs. self-hosting
The two paths look very different in terms of effort and cost.
API (easiest path)
At $0.28 per million output tokens, running agent loops through the API is accessible to seed-stage startups and solo developers. The 2,500 concurrency limit is generous. V4-Pro output pricing is $0.87 per million tokens, so you are paying about a third of that for a model that beats it on the agentic benchmarks DeepSeek published. For teams already building AI integrations, this is the obvious starting point.
Self-hosting (requires serious hardware)
The weights are MIT-licensed and ungated, so on-premise commercial use is completely unblocked. The catch is memory. Even though only 13B parameters activate per token, all experts stay resident in RAM. Unsloth’s dynamic GGUFs put a lossless 8-bit build at 162 GB and a 3-bit build at 103 GB, needing roughly 110 GB of combined RAM and VRAM at aggressive quantization. DeepSeek’s own vLLM example uses a single 4xGB300 node. Self-hosting suits mid-size or large enterprises with a serving cluster, or a single very well-specced workstation at 3-bit.
Architecture notes
For teams building on the model, a few details matter. The attention mechanism combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Conventional residual connections are replaced by Manifold-Constrained Hyper-Connections (mHC) with an expansion factor of 4 and 20 Sinkhorn-Knopp iterations. Pre-training used more than 32 trillion tokens with the Muon optimizer. Multi-token prediction depth is 1. There is no Jinja chat template; DeepSeek ships an encoding folder with encode_messages and parse_message_from_completion_text functions instead. For agentic use, DeepSeek recommends temperature 1.0, top_p 0.95, and up to 384K output tokens at high or max reasoning effort settings.
The DSpark paper reports 60 to 85 percent faster per-user generation versus the MTP-1 baseline at matched aggregate throughput. The DSpark vLLM flag for enabling it: --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'.
Our take
The DeepSWE jump from 7.3 to 54.4 is the number that jumps out. That is not a rounding error, it is a complete change in capability tier for software engineering agent tasks. If it holds under independent evaluation, V4-Flash-0731 becomes a serious option for any team running automated code review, PR generation, or repo-level agent workflows.
The pricing is genuinely competitive. At $0.28 per million output tokens with a 1M context window, you can run long multi-step agent conversations without the token cost becoming the bottleneck. That said, the harness caveat is real. Every number in the benchmark table came from DeepSeek using tooling it has not released. Before you redesign your workflow automation stack around these scores, run at least a small eval on tasks representative of your actual workload. Treat the published numbers as directionally useful, not definitive.
For teams already exploring where non-OpenAI models fit in the stack, this release is worth a direct test. The MIT license and ungated weights remove any deployment blocker for commercial on-premise use.
What to do about it
- Call the public beta API with a representative sample of your agent tasks and measure quality against your current model.
- Enable DSpark in vLLM if you are self-hosting and measure throughput gains on your actual query volume.
- Check the reasoning_effort parameter: use low for latency-sensitive tasks, high or max for complex multi-step agent runs, and cap output tokens at 384K for the latter.
- Wait for third-party harness evals before making V4-Flash-0731 your primary model in production pipelines that rely on code agent benchmarks.
If the DeepSWE and Cybergym numbers hold under independent testing, V4-Flash-0731 is the most cost-effective option for agentic coding workflows available today.
Frequently asked questions
What is the price of DeepSeek V4-Flash API?
DeepSeek lists deepseek-v4-flash at $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens, with a 2,500 concurrency limit.
How many parameters does DeepSeek V4-Flash-0731 have?
The base model has 284B parameters, with 13B activated per token. The full Hugging Face checkpoint including the DSpark speculative decoding draft module is 304B total.
How much memory do you need to self-host DeepSeek V4-Flash?
A lossless 8-bit quantization requires 162 GB, and a 3-bit build requires 103 GB of weight storage, needing roughly 110 GB of combined RAM and VRAM. DeepSeek's own example uses a single 4xGB300 node.
What changed between DeepSeek V4-Flash preview and V4-Flash-0731?
The architecture and model size are identical. The performance gains come entirely from re-post-training. The 0731 release also adds native Responses API support, Codex adaptation, and the DSpark speculative decoding module.