DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, as an open-weight model under the MIT License, and made it available on its API as deepseek-flash. DeepSeek describes it as a 552-billion-parameter mixture-of-experts model with native image understanding and a 1M-token context, using only 8 billion active parameters for input and 16 billion for output. On the hosted API it costs $0.30 per million input tokens and $1.20 per million output tokens at peak times, and half that off-peak.
DeepSeek's release announcement calls V4.1-Flash the smallest model in a new architecture family, built for higher capability, faster inference and scaling to larger models later. The model card gives the technical outline:
One number needs care. The model card describes 552B backbone parameters, while the Hugging Face file viewer reports 763B parameters in the stored weights. The card does not reconcile the two figures, so the 763B count should be treated as the storage figure to plan around and 552B as DeepSeek's description of the backbone.
The model card states that the repository and model weights are licensed under the MIT License, a short permissive license. That is unusually open for a model of this size. It allows commercial use, modification and redistribution with attribution, without the use-based restrictions found in some other model licenses. The weights are available now; this is not a promised future release.
The MIT License covers the weights and repository. It does not cover training data, and neither the announcement nor the model card describes a training-data release, so "open-weight under MIT" is a more precise description than "fully open source."
| Per million tokens (USD) | Peak | Off-peak |
|---|---|---|
| Input, cache hit | $0.006 | $0.003 |
| Input, cache miss | $0.30 | $0.15 |
| Output | $1.20 | $0.60 |
Source: DeepSeek Models & Pricing, checked October 10, 2026.
The pricing page lists peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Everything else is off-peak, including weekends. For teams in the Americas, much of the working day falls in off-peak hours.
The same page lists, for the model name deepseek-flash:
The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but are served by V4.1-Flash. DeepSeek's September 10 announcement said V4-Pro requests would also be routed to V4.1-Flash from September 14; the current changelog says DeepSeek will continue providing V4-Pro after September 14 with billing unchanged. The two records contradict each other. Check which model your requests actually reach if you depend on V4-Pro.
The changelog lists a long set of DeepSeek-reported scores, including 74.2 on DeepSWE v1.1, 31.2 on Terminal-Bench 4.0 and 90.9 on GPQA Diamond, and DeepSeek says V4.1-Flash outperforms its own larger V4-Pro. These are DeepSeek's measurements under its own harness and settings, and this article does not reproduce them. Comparing them with other vendors' published numbers is unreliable unless the harness and effort settings match.
Open weights are only useful if you can run them. Even with few active parameters per token, all experts must be held in memory, and the stored weights total hundreds of gigabytes in mixed FP8 and 8-bit formats. That points to multi-GPU servers rather than a single accelerator. DeepSeek says it is working with the open-source community on inference support and invites contact about large deployments.
For most teams, the realistic paths are the hosted DeepSeek API, a third-party inference provider serving the open weights, or self-hosting on dedicated GPU infrastructure where data control justifies the cost. Teams with data-residency or vendor-risk requirements should review where any hosted endpoint processes data before sending sensitive inputs. The open weights are what make self-hosting an option at all.
Teams that need open weights with permissive terms have few options at this capability level. The MIT License removes most legal friction for commercial use and fine-tuning.
Cost-sensitive API users can compare its $0.30 input and $1.20 output peak rates with the base rates of GPT-6 Luna and Claude Haiku 5.5, which both list at $0.10 input and $0.50 output. Those base rates have long-prompt conditions: Haiku 5.5 charges higher rates on requests whose prompts exceed 100,000 tokens, per Anthropic's pricing documentation, and Luna charges higher rates on requests above 272,000 input tokens, per OpenAI's model page; see the GPT-6 Luna and Claude Haiku 5.5 guides. At peak, DeepSeek costs more per token than those two at their base rates; off-peak, its rates halve to $0.15 input and $0.60 output, still above them. It reports results ahead of its own V4-Pro flagship, so the comparison that matters is cost per completed task.
Teams on V4-Flash are already being served by V4.1-Flash; switch the model name to deepseek-flash.
When was DeepSeek-V4.1-Flash released?
September 10, 2026.
Is DeepSeek-V4.1-Flash open source?
Its weights and repository are released under the MIT License, a permissive license. DeepSeek's announcement and model card do not describe a training-data release, so it is most accurately described as open-weight under a permissive license.
How much does the DeepSeek-V4.1-Flash API cost?
At peak: $0.30 per million input tokens on a cache miss, $0.006 on a cache hit, and $1.20 per million output tokens. Off-peak rates are half. Prices from DeepSeek's pricing page, checked October 10, 2026.
Discover more content:
