DeepSeek-V4-Flash is officially out of preview. The headline that really matters: a model with only 13B active parameters just beat the 49B-active V4-Pro preview on all nine agent benchmarks — purely through post-training, architecture untouched.
On July 31, 2026, DeepSeek-V4-Flash graduated to official status (DeepSeek-V4-Flash-0731), with public API access opening same day (source). Weights are MIT-licensed. Pricing is unchanged.
Same Architecture, Way Better Capability
Same CSA + HCA design. Same 284B total / 13B active parameters. Just one round of re-post-training. You can literally drop in the new weights — zero inference code changes.
| Metric | DeepSeek-V4-Flash-0731 |
|---|---|
| Total Parameters | 284B |
| Active Parameters | 13B |
| Architecture | CSA + HCA |
| Context Length | 1M |
| Max Output | 384K |
| Precision | FP4 + FP8 Mixed |
| New | DSpark speculative decoding |
| License | MIT |
DSpark is a new acceleration module — it speeds up decoding without touching anything in the output path.
Agent Benchmarks: 13B Bodies 49B
This isn't incremental. The post-training jump across all nine benchmarks is a full tier up (source):
| Benchmark | V4-Flash-0731 |
|---|---|
| Terminal Bench 2.1 | 82.7 |
| NL2Repo | 54.2 |
| Cybergym | 76.7 |
| DeepSWE | 54.4 |
| Toolathlon verified | 70.3 |
| Agent Last Exam | 25.2 |
| Automation Bench (Public) | 25.1 |
| DSBench-FullStack | 68.7 |
| DSBench-Hard | 59.6 |
DSBench is proprietary — use for directional comparison only, not cross-model absolute scores.
What's New
Responses API
Native support for OpenAI's Responses API — the stateful replacement for Chat Completions, built for agentic workloads. V4-Pro gets this in early August.
Codex Adaptation
Specifically tuned for Codex, the agent framework that powers Claude Code. If you're in the Claude Code + DeepSeek API camp, this is a real quality-of-life improvement. See the official config guide.
DSpark Speculative Decoding
Pure throughput play — doesn't change model output, doesn't add parameters, just makes tokens move faster.
Pricing and Concurrency
Exactly the same as preview:
| Item | Price (USD) | Price (CNY) |
|---|---|---|
| Cache-hit input | $0.0028 / MTok | ¥0.02 / MTok |
| Cache-miss input | $0.14 / MTok | ¥1 / MTok |
| Output | $0.28 / MTok | ¥2 / MTok |
Concurrency 2500. Thinking/non-thinking dual mode, JSON Output, Tool Calls — all there.
Worth noting: the official pricing page also carries a price-hike warning. We covered this in a separate piece, but the short version: get your caching strategy dialed in now.
Open Weights
MIT license, available on both HuggingFace and ModelScope:
- HuggingFace: deepseek-ai/DeepSeek-V4-Flash-0731
- ModelScope: deepseek-ai/DeepSeek-V4-Flash-0731
Base and Instruct variants. Fine-tune with Base, deploy with Instruct.
What to Watch
- Flash only this time — the Pro API, app, and web models are untouched.
- V4-Pro official is "coming soon," per DeepSeek's own words.
deepseek-chatanddeepseek-reasonerare fully retired (July 24, 15:59 UTC). They now route to v4-flash's non-thinking/thinking modes.- Don't sleep on the price-hike warning — calc your monthly spend now and lock in your caching strategy.