DeepSeek V4 Flash-0731: Official Release Brings Major Agentic and Coding Gains

On July 31, 2026, DeepSeek moved its V4-Flash model out of preview and into public beta with the release of DeepSeek-V4-Flash-0731. The update does not touch the model's size or architecture. Instead, DeepSeek re-post-trained the same 284-billion-parameter Mixture-of-Experts (MoE) design and unlocked noticeably stronger agent, coding, and tool-use behavior. For developers already using deepseek-v4-flash, the change is drop-in: keep the same model name and the API returns the new checkpoint.

This post breaks down what changed, how the benchmarks compare, what it costs, and how to deploy or self-host it.

What changed on July 31, 2026

The official API release keeps the existing endpoint and model name unchanged. Calling deepseek-v4-flash now routes to the 0731 checkpoint.

Key additions:

• Native Responses API format support, with adaptation for Codex-style agents.
• Tool calling and structured JSON output out of the box.
• Thinking and non-thinking modes selectable through the API.
• OpenAI-style and Anthropic-style interfaces remain supported.
• A 1-million-token context window is unchanged.

The model card and API docs are explicit: the architecture and size are identical to the April preview. All gains come from post-training, not a new design.

Architecture and size

The model ships on Hugging Face with the DSpark speculative decoding module attached, matching the structure of DeepSeek-V4-Flash-DSpark. The repo is reported at 304B parameters because the draft module sits on top of the 284B base.

Benchmarks: V4-Flash-0731 vs the preview

DeepSeek's published numbers show the official release pulling ahead of the earlier preview on almost every agentic and coding benchmark, and even surpassing the V4-Pro-Preview in several categories.

Two important caveats:

1. The code-agent tasks were run with the DeepSeek Harness minimal mode, which has not been publicly released. Third-party reproductions may differ.
2. DSBench-FullStack and DSBench-Hard are internal DeepSeek test sets, so they cannot be independently verified.

Still, the directional jump is large enough to matter: a model that was already cheap is now competitive with the larger preview on agentic work.

Pricing

DeepSeek's pricing makes V4-Flash attractive for high-volume agent loops:

That output price is roughly one-third of V4-Pro's $0.87. The concurrency limit is 2,500, which is enough for most seed-stage products and internal platform teams.

How to use it via the API

If you already call the DeepSeek API, change nothing except making sure your SDK or HTTP call uses deepseek-v4-flash:

1curl https://api.deepseek.com/chat/completions \
2  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
3  -H "Content-Type: application/json" \
4  -d '{
5    "model": "deepseek-v4-flash",
6    "messages": [{"role": "user", "content": "Write a Python function that validates an email address"}],
7    "temperature": 1.0,
8    "top_p": 0.95
9  }'

For agentic work, DeepSeek recommends:

• temperature = 1.0
• top_p = 0.95
• reasoning_effort set to low, high, or max
• Up to 384K output tokens at high and max effort

The model also supports the Responses API format and is adapted for Codex, so you can pass multi-turn tool calls and receive structured reasoning traces.

Self-hosting and deployment

Via API

The lowest-friction path. No GPU budget, no serving stack, and pricing is predictable.

Self-hosted

The weights are MIT-licensed and ungated, so on-premise or private-cloud commercial use is allowed. The challenge is hardware:

• Full precision / vLLM example: DeepSeek serves it on a single 4×GB300 node.
• Quantized via Unsloth: a lossless 8-bit GGUF build is around 162 GB; a 3-bit build is around 103 GB, needing roughly 110 GB of combined RAM plus VRAM.

Self-hosting is realistic for mid-size companies with a serving cluster, or a single well-specced workstation if you accept aggressive quantization. For startups and small teams, the hosted API is the practical default.

DSpark speculative decoding

If you self-host with vLLM, enable DSpark with one flag:

1--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

The DSpark paper reports 60–85% faster per-user generation on V4-Flash compared with the MTP-1 baseline at matched aggregate throughput.

Caveats and gotchas

• No Jinja chat template. DeepSeek ships an encoding/ folder with encode_messages and parse_message_from_completion_text helpers. Integrations that rely on a standard chat template will need minor adapter code.
• Vendor-reported benchmarks. The headline agentic numbers are from DeepSeek's own harness. Treat them as directional, not definitive, until independent reproductions appear.
• V4-Pro API unchanged. This update only affects deepseek-v4-flash. The V4-Pro API and the consumer app/web models are still on their previous checkpoints.

FAQ

Do I need to change my API call? No. deepseek-v4-flash now points to the 0731 checkpoint automatically.

Is the context window still 1M tokens? Yes.

Can I use it with Codex? Yes. The 0731 release is specifically adapted for Codex and supports the Responses API format.

Is self-hosting free? The weights are free and ungated, but you still pay for the hardware to run them.

When is V4-Pro officially releasing? DeepSeek says the official V4-Pro release will follow soon; the 0731 update is Flash-only.

Bottom line

DeepSeek-V4-Flash-0731 is a post-training upgrade that turns an already-cheap model into a stronger agent and coding workhorse. It keeps the same API contract, context window, and architecture, while adding native Responses API support, Codex adaptation, and benchmark numbers that beat the larger V4-Pro-Preview on several agentic tasks. For teams building code agents, autonomous tools, or high-volume API products, it is now one of the most cost-effective frontier options on the market.

Sources: DeepSeek API Docs — July 31, 2026 update, DeepSeek-V4-Flash-0731 model card on Hugging Face, MarkTechPost coverage. Benchmarks are vendor-reported.