Cette section n'existe pas encore en français — le contenu ci-dessous est en anglais. La traduction arrive.
DeepSeek V4 Flash
An incredibly cost-effective, open-weights reasoning model from China that rivals proprietary giants.
Quality
Modality
text
Context
1M tokens
Access
open weights
Fabian's Take
"It's almost suspiciously cheap. I use it through OpenRouter when I need to run a task hundreds of times and don't want to burn through API credits. The reasoning capability at this price point is completely disruptive."
DeepSeek V4 Flash is an open-weights model developed in China that has disrupted the pricing structure of the AI market. Built on an efficient Mixture-of-Experts (MoE) architecture, it offers reasoning capabilities that rival top-tier proprietary models from OpenAI and Anthropic, but at a fraction of the cost.
The pricing disruption
The API pricing for DeepSeek V4 Flash is so low it fundamentally changes what is possible in agentic workflows. Tasks that were previously too expensive to automate at scale — like running deep logical checks over thousands of documents — are suddenly viable. The model defaults to a “thinking” mode, enabling it to handle complex math and coding challenges effectively.
The Verdict
Best for: High-volume reasoning tasks where cost efficiency is critical.
Pros
- Extremely cheap API pricing (up to 100x cheaper than GPT-5.5)
- Default 'thinking' mode for strong logical reasoning
- Open-weights under an MIT license
Cons
- Training data and full codebase are not completely open
- Chinese-centric pretraining can occasionally influence nuances in other languages
Specs
- Pricing $0.14/M input, $0.28/M output (cache misses)
- Cost Tier free
- ⚡Speed Tier fast
- License MIT