DeepSeek V4 Flash 0731: Specs, Benchmarks, and API Guide
DeepSeek officially released DeepSeek-V4-Flash-0731 on July 31, 2026, replacing the earlier preview version of DeepSeek V4 Flash. The API release is currently in public beta, and developers do not need to change their existing integration: the API model identifier remains deepseek-v4-flash.
The release focuses primarily on stronger coding-agent and tool-use performance. According to DeepSeek, the model keeps the architecture and parameter scale of the preview version but receives a new post-training stage that substantially improves its agent capabilities.
Data note: Specifications, prices, and benchmark results in this article come from official DeepSeek sources as of August 2, 2026. Benchmark results are vendor-reported and should not be treated as fully independent evaluations.
Key specifications
| Item |
DeepSeek V4 Flash 0731 |
| Official checkpoint |
deepseek-ai/DeepSeek-V4-Flash-0731 |
| API model ID |
deepseek-v4-flash |
| Architecture |
Mixture of Experts, or MoE |
| Total parameters |
284 billion |
| Activated parameters |
13 billion per token |
| Context window |
1 million tokens |
| Maximum output |
384K tokens |
| Routed experts |
256 |
| Experts selected per token |
6 routed experts plus 1 shared expert |
| Reasoning modes |
Non-thinking, low, high, and max reasoning effort |
| API formats |
OpenAI-compatible, Anthropic-compatible, and Responses API |
| Model license |
MIT |
| Release status |
API public beta |
The distinction between total and activated parameters is important. DeepSeek V4 Flash contains 284 billion parameters, but its MoE router activates only a subset—approximately 13 billion parameters—for each token.
This makes it more compute-efficient than a dense 284B model, although it is still a very large model to download and self-host.
What changed in the 0731 release?
DeepSeek states that V4 Flash 0731 uses the same architecture and parameter scale as the preview checkpoint. The main change is additional post-training rather than a new foundation-model architecture.
The update introduces several practical improvements:
- Substantially stronger coding-agent results
- Better terminal operation and repository-level reasoning
- Improved tool selection and multi-step automation
- Native support for the Responses API
- Dedicated compatibility work for Codex-style coding agents
- Three reasoning-effort levels:
low, high, and max
- A DSpark speculative-decoding module for faster generation
This is a useful reminder that model quality is not determined only by parameter count. Better training data, reinforcement learning, tool-use training, and agent-oriented post-training can materially improve results without changing the underlying model size.
Architecture and long-context design
DeepSeek V4 Flash is part of the DeepSeek V4 family described in the paper DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence.
Its architecture includes several notable components.
Mixture-of-Experts routing
The repository configuration lists 256 routed experts, with six selected for each token, plus one shared expert. This selective activation is what allows the model to expose 284B total parameters while using roughly 13B activated parameters per token.
The design targets a balance between model capacity and inference cost.
Hybrid attention
DeepSeek V4 combines:
- Compressed Sparse Attention, or CSA
- Heavily Compressed Attention, or HCA
These mechanisms are intended to reduce the computation and memory required for extremely long prompts. The official configuration supports up to 1,048,576 positions, corresponding to the advertised one-million-token context window.
A large accepted context window does not automatically guarantee perfect retrieval across the entire prompt. Developers should still test long-document recall, instruction retention, and resistance to irrelevant information on their own workloads.
Manifold-Constrained Hyper-Connections
DeepSeek introduces Manifold-Constrained Hyper-Connections, abbreviated as mHC, to improve signal propagation and training stability across the network.
Training and precision
The V4 family was pretrained on more than 32 trillion tokens. For the released Flash checkpoint, MoE expert weights use FP4 precision, while most other parameters use FP8.
This mixed-precision format lowers memory and computation requirements, but local deployment remains an infrastructure-heavy task.
Official agent benchmark results
DeepSeek published the following comparison in the official DeepSeek V4 Flash 0731 model card.
| Benchmark |
V4 Flash 0731 |
Flash Preview |
V4 Pro Preview |
GLM-5.2 |
Opus-4.8 |
| Terminal Bench 2.1 |
82.7 |
61.8 |
72.1 |
81.0 |
85.0 |
| NL2Repo |
54.2 |
39.4 |
38.5 |
48.9 |
69.7 |
| Cybergym |
76.7 |
38.7 |
52.7 |
— |
83.1 |
| DeepSWE |
54.4 |
7.3 |
12.8 |
46.2 |
58.0 |
| Toolathlon-Verified |
70.3 |
49.7 |
55.9 |
59.9 |
76.2 |
| Agents' Last Exam |
25.2 |
15.8 |
16.5 |
23.8 |
25.7 |
| AutomationBench Public |
25.1 |
10.8 |
12.8 |
12.9 |
27.2 |
| DSBench-FullStack |
68.7 |
37.0 |
41.8 |
61.8 |
71.6 |
| DSBench-Hard |
59.6 |
25.8 |
31.1 |
54.5 |
71.7 |
Several improvements stand out:
- Terminal Bench increased from 61.8 to 82.7.
- Cybergym nearly doubled, rising from 38.7 to 76.7.
- DeepSWE increased from 7.3 to 54.4—roughly 7.5 times the preview score.
- Toolathlon-Verified improved by 20.6 points.
- The final Flash model outperformed the preview version of V4 Pro across every benchmark in this table.
The results also show that Flash 0731 is not automatically the strongest model in every comparison. Opus-4.8 remains ahead on all nine listed evaluations, although the gap is small on benchmarks such as Agents' Last Exam and AutomationBench.
How should these scores be interpreted?
The benchmark table is useful, but its methodology matters.
DeepSeek reports that public Code Agent tasks were evaluated with:
- DeepSeek Harness in its minimal mode
- The
max reasoning-effort setting
temperature = 1.0top_p = 0.95
At publication time, DeepSeek Harness had not yet been released. That makes exact third-party reproduction difficult.
Two benchmarks in the table also require additional caution:
- DSBench-FullStack is an internal full-stack development dataset.
- DSBench-Hard is an internal set of difficult coding-agent tasks.
Internal benchmarks can provide useful directional information, but they do not offer the same transparency as public datasets with accessible tasks, prompts, and scoring code.
The safest conclusion is that the 0731 post-training update appears to produce a large agent-performance improvement. Developers should still run application-specific evaluations before replacing an existing production model.
API availability and pricing
The public API continues to use the following model identifier:
deepseek-v4-flash
The version behind that identifier is currently DeepSeek-V4-Flash-0731.
As of August 2, 2026, official API pricing is:
| Billing item |
Price per 1M tokens |
| Input with cache hit |
$0.0028 |
| Input with cache miss |
$0.14 |
| Output |
$0.28 |
The API also lists:
- A one-million-token context window
- A maximum output length of 384K tokens
- JSON output support
- Tool calling
- Anthropic-compatible API support
- Responses API support
- A concurrency limit of 2,500 for V4 Flash
DeepSeek has announced a future peak-hour pricing policy. Once activated, prices during 09:00–12:00 and 14:00–18:00 Beijing time will be twice the regular price. The effective date remains subject to a separate official announcement, so production cost calculations should always use the current pricing page.
Calling DeepSeek V4 Flash through the API
The API is compatible with the OpenAI Chat Completions format.
curl https://caicaiapi.cloud/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
-d '{
"model": "deepseek-v4-flash",
"messages": [
{
"role": "system",
"content": "You are a careful coding assistant."
},
{
"role": "user",
"content": "Review this repository and propose a minimal fix."
}
],
"thinking": {
"type": "enabled"
},
"reasoning_effort": "high",
"stream": false
}'
For routine extraction, classification, and lightweight generation, non-thinking or low reasoning effort can reduce latency and token usage.
For repository-level coding, difficult debugging, or long-horizon agent tasks, the high or max settings may provide better results at the cost of longer responses and higher output-token consumption.
Responses API and coding-agent support
V4 Flash 0731 is currently the only DeepSeek V4 API model with official Responses API support. DeepSeek also states that the release has been specifically adapted for Codex.
This makes it especially relevant for:
- Repository-level coding assistants
- Terminal agents
- Automated code repair
- Multi-step tool execution
- CI failure analysis
- Codebase migration
- Full-stack feature implementation
Model quality is only one part of an agent system. Developers should also measure tool-call validity, recovery from failed commands, repeated-action behavior, test success rate, and whether the agent stops after completing its task.
Local deployment
The model weights are available under the MIT License, and the official documentation provides deployment examples for both vLLM and SGLang.
DSpark speculative decoding can be enabled in vLLM with a configuration similar to:
--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'
However, the official vLLM example uses a single node containing four GB300 GPUs. This should not be interpreted as a lightweight consumer-hardware model.
Although only 13B parameters are activated for each token, a self-hosted service still needs access to the complete checkpoint and its routing infrastructure. Most individual developers will find the hosted API more practical than local deployment.
Where does V4 Flash 0731 fit?
DeepSeek V4 Flash 0731 is best viewed as an efficiency-oriented agent model rather than a universally superior replacement for V4 Pro.
Its strongest potential use cases include:
- Coding and terminal agents
- Tool-driven workflow automation
- Long-document processing
- Repository analysis
- High-volume structured extraction
- Customer-facing AI applications where cost matters
- Routing tasks before escalating them to a larger model
The smaller activated parameter count and low API price make it attractive for production workloads. Meanwhile, its official agent scores suggest that post-training has compensated for some of the capability gap normally associated with a smaller active model.
For pure knowledge tasks, highly difficult reasoning, or workflows where the cost of failure is high, V4 Pro or another frontier model may still be the safer choice.
Limitations and open questions
Several points should remain part of any balanced evaluation:
- Most published 0731 scores currently come from DeepSeek itself.
- The DeepSeek Harness used in Code Agent tests was not publicly available at release time.
- Two prominently reported coding benchmarks are internal.
- Public-beta API behavior and pricing can still change.
- A one-million-token limit does not guarantee consistent use of every token.
- The 284B checkpoint remains expensive to self-host despite its 13B active parameter count.
- Strong benchmark performance does not guarantee reliable behavior on a specific production codebase.
A proper evaluation should compare models under the same prompts, tools, reasoning budgets, retry policies, and scoring rules.
Final assessment
DeepSeek V4 Flash 0731 is a meaningful update rather than a simple version-number change.
Its architecture remains the same as the preview release, but the new post-training stage produces major gains across coding-agent, terminal, cybersecurity, and automation benchmarks. With a one-million-token context window, 13B activated parameters, Responses API support, and aggressive API pricing, it is positioned as a practical model for high-volume agent workloads.
The official numbers are promising, especially when compared with DeepSeek V4 Pro Preview. They should still be treated as vendor-reported results until the evaluation harness and more independent tests become available.
For developers, the most sensible next step is not to rely on the leaderboard alone. Run V4 Flash 0731 against real repository tasks, measure successful outcomes rather than individual responses, and compare total cost—including retries and failed agent runs—against the alternatives.
Sources
