DeepSeek ships V4.1-Flash: MIT-licensed open-weight multimodal model with 1M-token context and sharply reduced KV cache

DeepSeek released V4.1-Flash, an open-weight text+image model with 1M-token context and an encoder-decoder design that cuts KV cache to roughly 1/4-1/8 of DeepSeek-V4-Flash. Sources disagree on size (552B vs 763B total parameters); it prices at $0.30/$1.20…

DeepSeek released V4.1-Flash, an open-weight multimodal model supporting text and image input with a 1M-token context, published on Hugging Face under an MIT license and also served via API. The two reports disagree on total parameters: The Decoder reports 552B, while Latent Space reports 763B using a novel causal encoder-decoder architecture. Both describe asymmetric active parameters - 8B per token for prefill and 16B for decode (Latent Space), summarized by The Decoder as 8-16B active parameters per token. The Decoder reports the model was trained from scratch on 45 trillion tokens of text and images. KV cache footprint is reduced to about a quarter of DeepSeek-V4-Flash in fast GPU memory and one-eighth when offloaded (The Decoder; Latent Space cites up to 1/8), aided by the encoder/decoder split and FP4 cache storage, and is 437x smaller per token than DeepSeek-V1. On benchmarks, The Decoder reports 74.2% on DeepSWE v1.1, narrowly beating Anthropic Opus 5 and OpenAI GPT-5.6 Sol; Latent Space reports an Artificial Analysis Intelligence Index score of 40 (above DeepSeek V4 Pro 0813) and a #1 open-weight ranking on the Vals Index ahead of Kimi K3. Latent Space gives API pricing of $0.30 per 1M input and $1.20 per 1M output tokens with a 50% off-peak discount; The Decoder notes it is served at V4-Flash prices. Baseten shipped day-0 support and Ollama began rolling the model out to paid subscribers. The reports frame the design differently: The Decoder attributes the gains to data and RL scaling rather than new algorithms, while Latent Space emphasizes the novel causal encoder-decoder architecture. The Decoder also notes remaining gaps versus closed frontier models on expert tasks and complex image understanding.

• Total parameters: 552B (The Decoder, 2026-09-10) vs 763B (Latent Space, 2026-09-12) - sources disagree
• Active parameters: 8B per token for prefill and 16B for decode (Latent Space); The Decoder cites 8-16B active parameters per token
• 1M-token context with text and image input
• Trained from scratch on 45 trillion tokens of text and images (The Decoder)
• KV cache: ~1/4 of DeepSeek-V4-Flash in fast GPU memory and 1/8 offloaded, and 437x smaller per token than DeepSeek-V1 (The Decoder); Latent Space cites up to 1/8 of V4 Flash
• Encoder/decoder split with FP4 cache storage (The Decoder); described by Latent Space as a novel causal encoder-decoder architecture
• Scores 74.2% on DeepSWE v1.1, narrowly beating Anthropic Opus 5 and OpenAI GPT-5.6 Sol (The Decoder)
• Artificial Analysis Intelligence Index score of 40, above DeepSeek V4 Pro 0813 (Latent Space)