DeepSeek v4.1 Flash
DeepSeekBalanced
ThinkingTool UseVisionStructured Output
About this model
DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on...
Performance Tier
Balanced
DeepSeek v4.1 Flash is a balanced model from DeepSeek : strong performance at a reasonable price.
Strong cost-performance ratio. Reliable for most professional use cases without premium pricing.
Pricing
This model is included in Elosia plansEco
Minimal cost. Ideal for very high volume or simple tasks.
| Type | per 1M tokens |
|---|---|
| Input (prompt) | $0.040 |
| Output (completion) | $0.144 |
| Cache read | $0.0016 |
Capabilities
Context Length1.0M
Max Output Tokens944K
TokenizerDeepSeek
Inputtext, image
Outputtext
Release DateSeptember 10, 2026
Benchmarks
General Intelligence
MMLU
Not reported
GPQA Diamond
90.9%
Mathematics
MATH-500
Not reported
Programming
HumanEval
Not reported
Reasoning
Humanity's Last Exam
39.2%
Multimodal
MMMU-Pro
77%
Agentic
Terminal-Bench 2.1
90.6%
Where does DeepSeek v4.1 Flash stand?
Compare its performance index against every other model.
Recommended Use Cases
CodingAnalysisResearchMathematicsData Extraction
Strengths
- Causal encoder-decoder MoE design that activates only 8B parameters per token during prefill and 16B during decode, keeping input-heavy agentic workloads cheap to serve
- Aggressive KV-cache compression (cross-layer reuse and FP4 caching) that keeps a 1M-token context lightweight in memory
- Native image understanding, trained jointly with text from the start of pre-training
- Continuously adjustable reasoning effort on a 1 to 100 scale to trade accuracy against inference cost
- MIT open weights with a published technical report and a 1M-token context
Limitations
- Weak factual recall: independent testing shows it almost always answers questions it does not know instead of abstaining
- Agentic scores depend heavily on the harness: the vendor's own table shows Terminal-Bench 2.1 moving by several points across agent scaffolds
- Verbose reasoning at maximum effort, which weighs on latency and token budget
- The Humanity's Last Exam and MMMU-Pro scores shown come from third-party testing; the vendor publishes MMMU-Pro only for the base model
Frequently asked questions
Resources
This model may use your data for training
