DeepSeek V4.1
#109 in Frontier LLMsDeepSeek · v4.1 · since 10. September 2026 · 3× · last seen Sep 13, 2026
20
Momentum
DeepSeek V4.1 was officially released as "DeepSeek-V4.1-Flash" – a separate "V4.1 Pro" does not yet exist. It is a multimodal Mixture-of-Experts model with 552B total parameters (8B active on input, 16B active on output), built on a new "Causal Encoder-Decoder" architecture with 20 encoder and 20 decoder layers. The model supports a 1-million-token context window, native image input, and is released as open weights under the MIT license on Hugging Face. It replaces the more expensive V4-Pro model in API routing starting September 14, 2026, and is priced significantly lower than comparable flagship models.
Momentum trend
15.06.13.09.
Features
| Key Benchmark (%) | HLE (Humanity's Last Exam): 36.8% full set / 39.1% text-only subset (official DeepSeek figure) |
| Context Window (Tokens) | 1,000,000 tokens (max output ~256,000+ tokens recommended) |
| License | MIT License (open model weights) |
| Multimodality | Native image understanding (text + image input, text output); supports JPEG, PNG, GIF, WebP |
| Platform | DeepSeek API (model ID: deepseek-flash), chat web/app, Hugging Face (weights for self-hosting via vLLM, SGLang, Transformers) |
| Price per 1M Tokens | Off-peak: $0.15 input / $0.60 output; Peak: $0.30 input / $1.20 output (cache hit: $0.003) |
| Release Date | September 10, 2026 (official release of DeepSeek-V4.1-Flash) |