# DeepSeek V4 Flash

DeepSeek, released 2026-07-19

**Verdict: recommended.** DeepSeek retired V4 Flash from its API on 10 September 2026: the deepseek-v4-flash id is still accepted, but its requests are served by V4.1-Flash and billed at the Flash rate of $0.15 input and $0.60 output per 1M tokens off-peak. The MIT-licensed weights (284B parameters, 13B active, 1M-token context) remain on Hugging Face, so this is now a frozen checkpoint for self-hosting and reproducible evals. For new API work, call deepseek-flash. Last verified 2026-07-24.

## Pricing

- Input: $0.15 / 1M tokens
- Output: $0.60 / 1M tokens
- Context: Retired Sep 10; id serves V4.1-Flash. Peak 2x M-F

## Strengths

- Self-hosted deployments of a frozen checkpoint: the MIT-licensed 284B-parameter (13B active) V4 Flash weights remain downloadable from Hugging Face, so pipelines already validated on V4 Flash can keep running it in-house
- Reproducible evals and audit trails: self-hosted V4 weights never change, while the deepseek-v4-flash API id began returning V4.1-Flash output on 10 September 2026 without any client-side change
- Classification, extraction and structured-output workloads on your own hardware, where SWE-bench Verified 79.0% capability comes with no per-token bill
- New API integrations: DeepSeek retired V4 Flash on 10 September 2026. The deepseek-v4-flash id now serves DeepSeek-V4.1-Flash, so calling it does not give you this model; use deepseek-flash directly
- Relying on the legacy API id for stable behaviour: DeepSeek says the old name is only temporarily routed to V4.1-Flash, and requests bill at the Flash rate of $0.15 input / $0.60 output per 1M off-peak, doubling 01:00-04:00 and 06:00-10:00 UTC Monday to Friday
- Hardest reasoning and multi-step agentic tasks: V4 Pro or Western flagships lead on sustained complex reasoning
- Vision and multimodal workloads: text-only model

## Benchmarks

- SWE-bench Verified: 79.0% (https://build.nvidia.com/deepseek-ai/deepseek-v4-pro/modelcard)
- SWE Pro (Resolved): 52.6% (https://build.nvidia.com/deepseek-ai/deepseek-v4-pro/modelcard)
- GPQA Diamond: 89.4% (https://www.requesty.ai/models/deepinfra/deepseek-ai-deepseek-v4-flash)
- τ²-Bench: 95.0% (https://www.requesty.ai/models/deepinfra/deepseek-ai-deepseek-v4-flash)
- Intelligence Index: 40.3 (https://designforonline.com/ai-models/deepseek-deepseek-v4-flash/)
- Coding Index: 56.2 (https://designforonline.com/ai-models/deepseek-deepseek-v4-flash/)

## Sources

- [AI Pricing Guru — DeepSeek API Pricing](https://www.aipricing.guru/deepseek-pricing/)
- [Design for Online — V4 Flash Review](https://designforonline.com/ai-models/deepseek-deepseek-v4-flash/)
- [Requesty — DeepSeek V4 Flash Specs & Benchmarks](https://www.requesty.ai/models/deepinfra/deepseek-ai-deepseek-v4-flash)
- [BenchLM: DeepSeek API Pricing: V4 Pro & Flash Rates](https://benchlm.ai/deepseek/api-pricing)
- [Tech Jacks Solutions — Running DeepSeek V4 Cost-Effectively](https://techjacksolutions.com/ai-tools/deepseek/running-deepseek-v4-cost-effectively/)
- [NVIDIA Build: deepseek-v4-pro Model Card](https://build.nvidia.com/deepseek-ai/deepseek-v4-pro/modelcard)
- [DeepSeek API Docs: Models and Pricing](https://api-docs.deepseek.com/quick_start/pricing/)
- [DeepSeek API Docs: DeepSeek-V4.1-Flash Release (2026/09/10)](https://api-docs.deepseek.com/news/news260910)
- [Hugging Face — deepseek-ai/DeepSeek-V4-Flash (MIT weights)](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash)

Canonical: https://neomanex.com/models/deepseek-v4-flash