DeepSeek-V4-Flash

On July 31, 2026, DeepSeek promoted DeepSeek-V4-Flash out of preview. The official build, tagged V4-Flash-0731, keeps the exact same 284B-parameter architecture we covered in our April V4 Preview breakdown and changes only one thing: the post-training. That one thing was apparently the whole story. On every one of the nine agent and coding benchmarks DeepSeek published, the retrained Flash now beats not just its own preview but V4-Pro-Preview, the 1.6T flagship it was supposed to sit under. Weights are on Hugging Face under MIT, API pricing stays at $0.14 in and $0.28 out per million tokens, and the model name is simply deepseek-v4-flash.

TL;DR

• What it is: the official release of DeepSeek's small V4 variant, a 284B Mixture-of-Experts model with 13B active parameters and a native 1M-token context, re-post-trained for agents and released under MIT on Hugging Face.
• Why it matters: on DeepSeek's own benchmark table, the 0731 build beats V4-Pro-Preview everywhere, and third-party Artificial Analysis puts it one point behind OpenAI's budget GPT-5.6 Luna at roughly 60% lower cost per task.
• The catch: 284B total parameters still means server-class hardware for self-hosting, the preview's high hallucination rate has no published 0731 re-measurement yet, and the API is labeled public beta.

What Actually Landed in DeepSeek-V4-Flash-0731

The official changelog is unusually direct about what changed: nothing in the architecture. "DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained." Same hybrid CSA plus HCA attention, same FP4 expert weights, same 1M context. If you want the architecture tour, read the April article; all of it still applies.

Three things are genuinely new:

• A full agentic retraining pass. The gains below came entirely from post-training, which is a quiet argument that open-weight labs still have large headroom without touching pretraining budgets.
• Native Responses API support. The changelog notes the model "natively supports the Responses API format and is specifically adapted for Codex," so it drops into OpenAI-shaped agent harnesses without adapter glue.
• The DSpark checkpoint. The 0731 release ships with a speculative-decoding module attached, which is also why the Hugging Face parameter counter reads 304B instead of 284B: the extra ~20B is the draft module, not the model growing.

Benchmarks: The Small Model Ate the Big One

All numbers below are DeepSeek-reported, from the 0731 model card. We have not run these ourselves.

Read the DeepSWE row twice. The preview scored 7.3. The same architecture, after retraining, scores 54.4. The changelog adds four more agent evals in the same pattern: DSBench-FullStack 68.7, DSBench-Hard 59.6, Toolathlon verified 70.3, and the still-hard tail of Agent Last Exam at 25.2 and Automation Bench (Public) at 25.1. Across all nine published benchmarks the 0731 build lands above V4-Pro-Preview, which makes the current DeepSeek lineup mildly absurd: the cheap model is the good one until V4-Pro gets the same retraining pass.

Third-party numbers point the same direction. The Decoder reports an Artificial Analysis Intelligence Index of 50 versus 51 for GPT-5.6 Luna, "about 60 percent less per task, even after OpenAI's 80 percent price cut," a GDPval jump from 1,189 to 1,559 Elo, and 12% fewer tokens used than the preview. For context, when we covered the preview in April, V4-Pro led the open-weights GDPval board at 1,554. The Flash just walked past that too.

Pricing: Still the Cheapest Seat at the Table

API pricing is unchanged from the preview, per the DeepSeek platform docs:

The cache-hit price is a 98% discount, which matters more than it looks for agents: long-horizon loops re-read the same context constantly, so a well-structured agent pays the cache-hit rate for most of its input tokens. Combine that with the 1M-token window and the KV-cache compression from the V4 architecture, and this stays the cheapest way to run serious long-context agent workloads on someone else's GPUs.

What It Takes to Run It Yourself

Honesty section: 13B active parameters does not mean 13B-sized hardware. The full FP4 plus FP8 checkpoint still wants a single big-memory accelerator or a small tensor-parallel group.

The vLLM single-GPU config from the official recipe is a useful reference for the moving parts (FP8 KV cache, the V4 tokenizer mode, and MTP speculative decoding at 3 draft tokens):

vllm serve deepseek-ai/DeepSeek-V4-Flash \
  --tensor-parallel-size 1 --kv-cache-dtype fp8 \
  --trust-remote-code --gpu-memory-utilization 0.92 \
  --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice --reasoning-parser deepseek_v4 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Hands-On: Point Your Agent at It in Under 10 Minutes

The API is OpenAI-compatible, so switching an existing agent stack is a one-line change:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DEEPSEEK_API_KEY",
    base_url="https://api.deepseek.com/v1",
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[
        {"role": "user", "content": "Read this repo and propose a fix plan."},
    ],
)
print(response.choices[0].message.content)

For self-hosting, grab the weights and follow the vLLM recipe above:

pip install -U "huggingface_hub[cli]"
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash-0731

If you are on the Codex or Responses API side of the fence, the 0731 build is explicitly tuned for that harness, which is the first time DeepSeek has shipped a model adapted for a specific third-party agent frontend.

Limitations and Gotchas

• Hallucination is an open question. The preview measured a 96% hallucination rate on Artificial Analysis's AA-Omniscience eval, among the worst of its peers. Nobody has published a 0731 re-measurement yet. Until someone does, keep retrieval grounding in fact-heavy pipelines.
• No Jinja chat template. Prompt encoding still goes through DeepSeek's Python encoder scripts in the repo. Day-one preview users tripped on this; nothing changed.
• The agent ceiling is real. Beating V4-Pro-Preview is impressive, but Agent Last Exam at 25.2 and Automation Bench at 25.1 are absolute scores. Long-horizon autonomy is still mostly unsolved, for everyone.
• Param-count confusion. Hugging Face shows 304B for the 0731 repo because the DSpark speculative-decoding module ships attached. The model itself is unchanged at 284B.
• Public beta API. DeepSeek labels the 0731 API a public beta, so expect endpoint behavior to shift.

Who Should Use It

If you run coding or tool-use agents against a hosted API and your bill matters, this is now the default open-weight choice: near-Luna scores at 40% of the price, MIT weights as your exit hatch if the API terms ever change. That exit hatch is not theoretical; we made the case for open-weight fallbacks when a US export order switched off a closed frontier model overnight. If you self-host, V4-Flash on a single H200-class card with quantization remains the practical play, same as the preview. And if you were waiting on the V3 line to get a proper successor for agent work, this is it.

What to watch next: a V4-Pro build with the same post-training recipe. If retraining alone took Flash from 7.3 to 54.4 on DeepSWE, the same pass over the 1.6T model is the obvious next shoe to drop.

Sources and Further Reading

• deepseek-ai/DeepSeek-V4-Flash-0731 on Hugging Face (weights, benchmark table)
• DeepSeek-V4-Flash-DSpark (speculative-decoding checkpoint)
• DeepSeek official release announcement (V4-Flash-0731)
• DeepSeek API changelog (official release notes and benchmark list)
• DeepSeek-V4 technical report (arXiv)
• vLLM serving recipe for DeepSeek-V4-Flash
• Artificial Analysis: DeepSeek V4 Flash 0731 analysis
• The Decoder: Flash matches GPT-5.6 Luna at 60% lower cost
• SingularityByte: DeepSeek-V4 Preview breakdown (April launch, full architecture tour)

Tested on: not independently benchmarked. Every score in this article is DeepSeek-reported or third-party (Artificial Analysis, The Decoder) as labeled; we have not run the 0731 build locally.
Date checked: 2026-08-01