DeepSeek-V4-Flash-0731 was DeepSeek's July 2026 agentic re-post-train of its Flash model, scoring 54.4 on DeepSWE. DeepSeek has since retired it on the API, so new projects should call deepseek-flash, which serves V4.1-Flash-1017. The MIT-licensed weights stay on Hugging Face for self-hosting and for reproducing older results.
What changed
DeepSeek-V4-Flash-0731, released July 31, 2026, is a retired DeepSeek model. DeepSeek's API docs say V4 Flash was retired on September 10, 2026 and the old names now route to V4.1-Flash-1017. On DeepSeek's own model card it scored 82.7 on Terminal Bench 2.1 with a 1,048,576-token context window.
Where it sits
- $0.262/M$ per 1M tokensBlended price (3:1)Lower is better#19 / 91peer median $1.69/Mvendor price, checked by HokAI
- 149 tok/stokens/sOutput speedHigher is better#16 / 54peer median 93 tok/scited: Artificial Analysis
- 88.1%% correctGPQA DiamondHigher is better#30 / 60peer median 87.9%per source, see benchmark scores
Cheaper than 77% of the 91 GA models with a published price, and mid-pack on GPQA Diamond (rank 30 of 60). Ranked against GA models; this record is not GA.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: DeepSeek · Family: DeepSeek V4
Context window: 1,048,576 tokens · Max output: 65,536
Input modalities: text, tool-calls · Output: text, tool-calls
About DeepSeek-V4-Flash-0731
DeepSeek-V4-Flash-0731 is the July 31, 2026 re-post-trained build of DeepSeek's efficiency-tier Flash model, and DeepSeek no longer serves it. DeepSeek retired the V4 Flash line when it released DeepSeek-V4.1-Flash on September 10, 2026. Its API docs now say the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but requests are served by DeepSeek-V4.1-Flash-1017, the October 17, 2026 update, and billed at the Flash price.
While it was live, the 0731 build was the official release of V4 Flash, superseding the April preview of the DeepSeek-V4 family. DeepSeek's model card on Hugging Face reports Terminal Bench 2.1 at 82.7 (preview 61.8) and DeepSWE at 54.4 (preview 7.3), and says it outperformed the preview of DeepSeek-V4-Pro on those tests. The weights are MIT licensed, and the card lists a speculative decoding module attached to the checkpoint. The Hugging Face page shows 304 billion parameters; MarkTechPost's launch coverage reported 284 billion total with 13 billion active per token. The context window is 1,048,576 tokens.
What is still useful on this page: the weights stay on Hugging Face, so teams can self-host the build or reproduce results from July to September 2026. Hosts such as OpenRouter and Fireworks AI had per-model listings for it, so check each provider before assuming which weights a request reaches. Agent tools such as Claude Code can be pointed at DeepSeek's Anthropic-format endpoint, which now reaches the successor.
Who should not pick this model: anyone starting new API work. Call the deepseek-flash alias instead, which adds image input and runs at the current Flash rates, and see the cheapest LLM API guide for how those rates compare. The benchmark figures are also DeepSeek's own. In the same model card table, the closed model Opus-4.8 scores 85.0 on Terminal Bench 2.1 and 58.0 on DeepSWE, ahead of the 0731 build on both, and GLM-5.2 is listed at 81.0 and 46.2.
Pricing
DeepSeek no longer lists a separate rate for this build. The legacy V4 Flash names are billed at the current Flash price: $0.15 per 1M input tokens (cache miss) and $0.60 per 1M output tokens off-peak, $0.30 and $1.20 at peak, with cache hits at $0.003 off-peak and $0.006 at peak. Peak runs on weekday mornings UTC, and off-peak rates are half of peak (per the DeepSeek pricing page). The [cheapest LLM API guide](/guides/cheapest-llm-api-2026) compares current rates.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0045 | $0.0006 | $0.0051 |
| Support reply | $0.0003 | $0.0002 | $0.0005 |
| One coding agent run | $0.030 | $0.012 | $0.042 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Retired on the DeepSeek API: DeepSeek retired the V4 Flash line on September 10, 2026; the old model names are routed to DeepSeek-V4.1-Flash-1017.
- 1M-Token Context Window: A 1,048,576-token context window held a full repository or document set in one request.
- Agentic Re-Post-Train: The 0731 build re-trained the April preview for tool use, lifting Terminal Bench 2.1 from 61.8 to 82.7 and DeepSWE from 7.3 to 54.4 on DeepSeek's model card.
- MIT-Licensed Open Weights: Open MIT-licensed checkpoint in FP8 and FP4 formats, with community GGUF and MLX builds for local runs.
- Sparse MoE Design: A Mixture-of-Experts model; the Hugging Face page lists 304B parameters, while launch coverage reported 284B total with 13B active per token.
Pros
- DeepSWE rose several-fold between the April preview and this build on DeepSeek's model card, with the same architecture.
- MIT-licensed weights stay downloadable, so self-hosting teams keep a permissive-license path after the API retirement.
- The very long context window supports whole-repository prompts.
Cons
- DeepSeek retired it on its own API in September 2026; the old model name now returns the successor model instead.
- No published system card or disclosed red-team partners for this specific build.
- Vision input support was reported inconsistently across providers; the successor model lists image input.
- Benchmark figures come from DeepSeek's own card and a launch article, not an independent re-run.
Benchmarks
- MMLU: 89% vendor-reported · 26 Oct 2026 — General-knowledge exam across 57 subjects, % correct.
- DeepSWE: 54.4 vendor-reported · 26 Oct 2026
- MMLU-Pro: 86.4% independent · 01 Aug 2026 — A harder version of the 57-subject knowledge exam, % correct.
- HumanEval: 90% vendor-reported · 26 Oct 2026 — Small programs that must pass hidden tests, % passing.
- GPQA Diamond: 88.1% independent · 01 Aug 2026 — PhD-level science questions that are hard to search for, % correct.
- Terminal-Bench 2.1: 82.7% vendor-reported · 26 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
- AA Intelligence Index: 34 independent · 26 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- Output speed: 149 tok/s independent · 26 Oct 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
Is DeepSeek-V4-Flash-0731 still available on the DeepSeek API?
No. DeepSeek ended API service for this build with the September 10, 2026 launch of V4.1 Flash. Calls that use the legacy model names still succeed, but the reply comes from the successor, so outputs and benchmark behavior differ from the July checkpoint.
What replaced DeepSeek-V4-Flash-0731?
DeepSeek-V4.1-Flash-1017, released October 17, 2026, is the model behind the deepseek-flash name. DeepSeek's changelog says the calling method and pricing are unchanged from V4.1 Flash, and the V4.1 line adds native image input that the 0731 build was not confirmed to have.
What does the old DeepSeek V4 Flash model name cost now?
The old name is charged on the V4.1 Flash rate card, which is 15 cents in and 60 cents out per million tokens off-peak. Weekday mornings in UTC count as peak and cost double. A cache hit is billed at $0.003 per million tokens off-peak.
Can I still run DeepSeek-V4-Flash-0731 on my own hardware?
Yes. The weights remain on Hugging Face under the MIT license, so the exact checkpoint can be self-hosted. DeepSeek's card recommends temperature 1.0 and top_p 0.95 for agentic use, and a maximum output length of 384K tokens at the high and max reasoning levels.
How did DeepSeek-V4-Flash-0731 score on agentic coding tests?
DeepSeek's own model card lists 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE, up from 61.8 and 7.3 for the April preview. The same table gives Opus-4.8 85.0 and 58.0, and GLM-5.2 81.0 and 46.2, so these are vendor-reported figures rather than an independent re-run.
HokAI guides that mention DeepSeek-V4-Flash-0731
- Fireworks AI vs Together AI (2026): Managed Platform or Neocloud?: Fireworks AI and Together AI charge the same per token on seven shared open models. So pick by setup: Together rents GPU clusters, Fireworks sells managed ones.
- The AI Tool Ecosystem in 2026: Buy the Meter, Not the Category: Four meters set your AI infrastructure bill: tokens, machine seconds, vendor credits, seats. Modal, Replicate, E2B and Together AI re-priced 26 October 2026.