Page last updated on
DeepSeek-V4-Flash-0731
By DeepSeek
The official DeepSeek-V4-Flash build, retrained for agents and coding. A 284B MoE with 13B active parameters and 1M context, it beats DeepSeek-V4-Pro-Preview on every agent benchmark DeepSeek published.
Specifications
- Total parameters
- 284B (13B active)
- Active parameters
- 13B
- Architecture
- MoE
- Architecture class
- DeepseekV4ForCausalLM
- Attention
- Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA)
- Context window
- 1M
- Vocab size
- 129,280
- Modality
- Text
- Precision
- FP4 + FP8 Mixed
- License
- MIT
- Released
- July 2026
- Recommended hardware
- 8× H2008× MI325X
- Best for
- Cost-sensitive coding agents, tool use, and long-context automation
Good to know
-
DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the April 2026 preview. It entered public beta on the API on July 31, 2026. The architecture and the parameter count are unchanged from the preview. Every gain comes from a redone post-training run.
-
This build natively supports the Responses API format and was adapted for Codex, the OpenAI coding assistant that ships as a CLI, a ChatGPT desktop app, and a VS Code extension. You can register DeepSeek as a model provider directly, instead of routing through a compatibility layer.
-
The larger sibling caught up on August 12, 2026. DeepSeek-V4-Pro-0813 is the official V4-Pro build, and it beats this model on every benchmark DeepSeek published. V4-Pro also costs three times more per output token under the pricing that took effect on August 16, 2026.
Architecture
DeepSeek-V4-Flash-0731 is the same model as the April preview. The weights were re-post-trained, nothing else changed, and that is the interesting part.
The architecture is the DeepSeek-V4 design at the smaller size: a 284B-parameter Mixture-of-Experts that activates 13B parameters per token, across 43 layers with 256 routed experts, 1 shared expert, and 6 experts selected per token. Attention interleaves Compressed Sparse Attention (CSA) with Heavily Compressed Attention (HCA). CSA compresses blocks of the KV cache and attends to a few relevant entries. HCA folds much larger spans into single entries. Together they hold a 1,048,576-token window without paying dense-attention memory for it.
Two details are specific to this checkpoint:
- DSpark speculative decoding ships attached. The released weights include a speculative decoding module targeting layers 40, 41, and 42, enabled with a single flag.
- Selectable reasoning effort. The
reasoning_effortparameter takes low, high, or max. The published scores use the higher settings, and DeepSeek recommends allowing up to 384K output tokens there.
For more models like this one, browse the full open source LLM directory or read the open source LLM ecosystem statistics.
Benchmarks
DeepSeek published nine agent and coding benchmarks for the 0731 build. The headline result is that the cheap model passes the expensive one: 0731 beats DeepSeek-V4-Pro-Preview on all nine, despite activating 13B parameters against 49B. Terminal Bench 2.1 climbs to 82.7 and DeepSWE to 54.4, up from 7.3 on the preview. The gap to Claude Opus 4.8 narrows to under three points on Terminal Bench 2.1 and Agents' Last Exam, though Opus 4.8 keeps a clear lead on NL2Repo and DSBench-Hard.
DS-V4-Flash-0731 vs frontier models
Scores published by DeepSeek in the V4-Flash-0731 model card. Higher is better on every benchmark shown. Scores marked n/a were not reported. V4-Flash Preview and V4-Pro Preview are the April 2026 builds.
Each model keeps the same color across every benchmark, stacked top to bottom in this order. Hover or focus a row to read all six scores.
- Opus-4.8
- GLM-5.2
- V4-Flash Preview
- V4-Pro Preview
- DS-V4-Flash-0731
View all scores as a table
| Benchmark | Opus-4.8 | GLM-5.2 | V4-Flash Preview | V4-Pro Preview | DS-V4-Flash-0731 |
|---|---|---|---|---|---|
| Agentic & Coding | |||||
| Terminal Bench 2.1 (Acc) | 85.0 | 81.0 | 61.8 | 72.1 | 82.7 |
| NL2Repo (Pass@1) | 69.7 | 48.9 | 39.4 | 38.5 | 54.2 |
| Cybergym (Pass@1) | 83.1 | n/a | 38.7 | 52.7 | 76.7 |
| DeepSWE (Resolved) | 58.0 | 46.2 | 7.3 | 12.8 | 54.4 |
| Toolathlon-Verified (Pass@1) | 76.2 | 59.9 | 49.7 | 55.9 | 70.3 |
| Agents' Last Exam (Pass@1) | 25.7 | 23.8 | 15.8 | 16.5 | 25.2 |
| AutomationBench Public (Pass@1) | 27.2 | 12.9 | 10.8 | 12.8 | 25.1 |
| DSBench-FullStack (Pass@1) | 71.6 | 61.8 | 37.0 | 41.8 | 68.7 |
| DSBench-Hard (Pass@1) | 71.7 | 54.5 | 25.8 | 31.1 | 59.6 |
Source: the DS-V4-Flash-0731 model card. Best score in each row is marked.
Third-party evaluations
Independent evaluations landed within days of the open-weights release. They agree on the shape of the result: the 0731 build is a large jump over the April preview, it sits in the top tier of open models, and the price puts it on the cost-performance frontier.
API pricing
The 0731 build shipped without a price change. It is served on the existing deepseek-v4-flash endpoint at $0.14 per million cache-miss input tokens and $0.28 per million output tokens, so applications already on that endpoint picked up the agent gains without touching a line of code.
| Model | Input · cache hit | Input · cache miss | Output |
|---|---|---|---|
| DeepSeek-V4-Flash-0731 | $0.0028 | $0.14 | $0.28 |
| DeepSeek-V4-Pro | $0.003625 | $0.435 | $0.87 |
Per 1M tokens, as of August 6, 2026. Official pricing.
Against the closed frontier, output at $0.28 per million tokens is roughly 99% below Claude Opus 4.8 at $25 and about 99% below GPT-5.6 Sol at $30. The comparison that matters more is the near tier: GPT-5.6 Luna scores one point higher on the Artificial Analysis Intelligence Index and lists at $1.20 per million output tokens, so V4-Flash-0731 delivers comparable intelligence at roughly a quarter of the output price.
The table above shows the rates that applied through August 15, 2026. They are no longer current. At 16:00 UTC on August 16, 2026, alongside the official DeepSeek-V4-Pro-0813 release, DeepSeek replaced the flat rate with a peak and off-peak structure. V4-Flash output went to $1.32 per million tokens at peak and $0.66 off-peak, against $0.28 before. Cache-miss input went to $0.44 at peak and $0.22 off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC.
DeepSeek gave no reason for the increase. One independent read comes from dax of OpenCode, as they reproduced the current DeepSeek prices on rented GPUs. On that basis they argued the increase is traffic shaping under load, not a sign that DeepSeek is losing money at these rates.
Sources: DeepSeek pricing · Claude pricing · OpenAI pricing · Independent read on the price increase
What people are saying
“Mad respect to DeepSeek... It was the first time many people realized that open-weight models could genuinely compete with frontier models. Now again, DeepSeek has dropped V4 Flash. No hype. No fake marketing. No unnecessary drama. Just a model that competes with the best frontier models while being 89 times cheaper. Crazy.”
Pratham
@Prathkum · APILayer
X Aug 2, 2026
“With a score of 50 on the Artificial Analysis Intelligence Index, it lands among the top 3 open weights models on the leaderboard. The weights are released under the MIT license, allowing unrestricted commercial use and modification.”
Artificial Analysis
@ArtificialAnlys
X Jul 31, 2026
“Been playing with DeepSeek V4 flash 0731 all day and that's how much i spent, i don't know what to tell you but i always use subs because API cost a lot but for the first time i get really great performance for almost free and the Cache hits👌.”
Elshayib
@elshayib_ · Full-stack developer
X Jul 31, 2026
“现在全世界大部分模型都进入了DeepSeek斩杀区，能力差、价格贵的只能等着被斩杀了🫡 (Most models in the world have now entered the DeepSeek kill zone. The ones that are weak and expensive can only wait to be cut down.)”
Max For AI
@MaxForAI · Head of growth at LobeHub
X Aug 1, 2026
“on the upcoming deepseek price increase we've been able to reproduce their current prices even on rented GPUs so this likely isn't because they're "losing money" it's traffic shaping because they are overloaded”
dax
@thdxr · OpenCode
X Aug 6, 2026
“Last night I gave DS4 Flash and 5.6 Sol the same exact task - asked to go study up on how some things in the repo work, and then make a code change... DS4F’s result was practically the same as Sol, which surprised me. It was also ~34x times cheaper. if you have the hardware for it you can even run this locally which is nuts. Possibly maybe a new era?”
yags
@yagilb · LM Studio
X Aug 2, 2026
“Mad respect to DeepSeek... It was the first time many people realized that open-weight models could genuinely compete with frontier models. Now again, DeepSeek has dropped V4 Flash. No hype. No fake marketing. No unnecessary drama. Just a model that competes with the best frontier models while being 89 times cheaper. Crazy.”
Pratham
@Prathkum · APILayer
X Aug 2, 2026
“With a score of 50 on the Artificial Analysis Intelligence Index, it lands among the top 3 open weights models on the leaderboard. The weights are released under the MIT license, allowing unrestricted commercial use and modification.”
Artificial Analysis
@ArtificialAnlys
X Jul 31, 2026
“Been playing with DeepSeek V4 flash 0731 all day and that's how much i spent, i don't know what to tell you but i always use subs because API cost a lot but for the first time i get really great performance for almost free and the Cache hits👌.”
Elshayib
@elshayib_ · Full-stack developer
X Jul 31, 2026
“现在全世界大部分模型都进入了DeepSeek斩杀区，能力差、价格贵的只能等着被斩杀了🫡 (Most models in the world have now entered the DeepSeek kill zone. The ones that are weak and expensive can only wait to be cut down.)”
Max For AI
@MaxForAI · Head of growth at LobeHub
X Aug 1, 2026
“on the upcoming deepseek price increase we've been able to reproduce their current prices even on rented GPUs so this likely isn't because they're "losing money" it's traffic shaping because they are overloaded”
dax
@thdxr · OpenCode
X Aug 6, 2026
“Last night I gave DS4 Flash and 5.6 Sol the same exact task - asked to go study up on how some things in the repo work, and then make a code change... DS4F’s result was practically the same as Sol, which surprised me. It was also ~34x times cheaper. if you have the hardware for it you can even run this locally which is nuts. Possibly maybe a new era?”
yags
@yagilb · LM Studio
X Aug 2, 2026
Anecdotes
One prompt, 32 minutes, $0.07
Developer Elshayib (@elshayib_) ran DeepSeek-V4-Flash-0731 inside Hermes Agent from a single prompt. The agent worked for 32 minutes and the run cost $0.07. As they put it, the model is cheap enough that two dollars of API credit can last a full day.
A Beijing bar that serves API tokens with the Wi-Fi
Max For AI (@MaxForAI) reported that AGI Bar in Beijing began handing out DeepSeek-V4-Flash API access as a perk. Customers who connect to the in-store Wi-Fi get the API free, with no token cap while they are there. There is no subscription and no signup. You paste the base URL and the API key into Claude Code, Codex, Cursor, Cherry Studio, or any other client that speaks the OpenAI API format. Max For AI framed it as a change in defaults: inference is starting to look less like a metered subscription and more like water, power, or Wi-Fi that a venue simply provides.
Run it yourself
vllm serve deepseek-ai/DeepSeek-V4-Flash-0731
sglang serve --model-path deepseek-ai/DeepSeek-V4-Flash-0731Frequently asked questions
Is DeepSeek-V4-Flash-0731 open source?
Yes. The weights ship under the MIT license on Hugging Face, which allows unrestricted commercial use, modification, and redistribution.
What is new in the 0731 build compared to DeepSeek-V4-Flash Preview?
Only the post-training. The architecture, the 284B total and 13B activated parameter counts, the 1M context window, and the API pricing are all unchanged. The retraining targeted coding, agents, reasoning, and tool use. Terminal Bench 2.1 went from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4.
Is DeepSeek-V4-Flash-0731 better than DeepSeek-V4-Pro?
On the nine agent and coding benchmarks DeepSeek published, yes. The 0731 build beats V4-Pro-Preview on all nine, including 82.7 against 72.1 on Terminal Bench 2.1 and 54.4 against 12.8 on DeepSWE. The caveat is timing. Those numbers compare a freshly post-trained official build against a three-month-old preview, so the fair comparison arrives when the official V4-Pro ships.
How much does DeepSeek-V4-Flash-0731 cost?
Through the DeepSeek API it costs $0.14 per million cache-miss input tokens, $0.0028 per million cache-hit input tokens, and $0.28 per million output tokens. The 0731 release did not change these rates.
Does DeepSeek-V4-Flash-0731 work with Codex?
Yes. This build natively supports the Responses API format and was adapted for Codex, so you can add DeepSeek as a model provider rather than routing requests through a compatibility layer. Codex ships as a CLI, a ChatGPT desktop app, and a VS Code extension.
Keep exploring
2026-08-13
- Parameters
- 1.6T (49B active)
- Architecture
- MoE
- Context window
- 1M
- Modality
- Text
- License
- MIT
- Recommended Hardware
- 8× H2008× B200
View full details