DeepSeek releases open-weight DeepSeek-V4.1-Flash: 1M-token-context multimodal MoE with KV cache cut to 890 bytes/token

DeepSeek shipped DeepSeek-V4.1-Flash, an MIT-licensed multimodal Mixture-of-Experts model with a 1M-token context that activates 8B parameters at prefill and 16B at decode and cuts the global KV cache to 890 bytes per token (~1/4 of DeepSeek-V4-Flash, 437x…

DeepSeek released DeepSeek-V4.1-Flash, an open-weight, MIT-licensed multimodal Mixture-of-Experts model that was trending #28 on Hugging Face as of 2026-09-10. It uses a novel causal encoder-decoder architecture with Compressed Sparse Attention 2 (CSA2) and FP4 (E2M1) KV-cache quantization, reducing the global KV cache to 890 bytes per token — about one quarter of DeepSeek-V4-Flash in fast GPU memory and one-eighth when offloaded, and 437x smaller per token than DeepSeek-V1. Sources disagree on total size: Reports 1-2 describe a 552B-parameter backbone plus a separate 196B Engram conditional-memory module, Report 3 calls it 552B total, and Report 4 (Latent Space) calls it 763B total; all agree on 8B active parameters per token at prefill and 16B at decode. The model was trained from scratch on 45T multimodal tokens (text and images), with context extended to 1M tokens after sparse-attention training at 64K sequence length. Post-training used SFT, RL, and on-policy distillation with large-scale automated synthesis of agentic tasks and a controllable reasoning-effort setting from 1 to 100. Reported benchmark results: 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1, narrowly beating Anthropic Opus 5 and OpenAI GPT-5.6 Sol on agent benchmarks; Artificial Analysis scored it 40 on its Intelligence Index, above DeepSeek V4 Pro 0813, and Vals ranked it the #1 open-weight model ahead of Kimi K3. The Decoder attributes the gains to data and RL scaling rather than new algorithms and notes gaps remain versus closed frontier models on expert tasks and complex image understanding. Weights are on Hugging Face under MIT with vLLM and SGLang support, plus Baseten day-0 support and an Ollama rollout to paid subscribers; it is also served via API at $0.30 per 1M input and $1.20 per 1M output tokens with a 50% off-peak discount (Report 4; Report 3 says V4-Flash prices).

• Model: DeepSeek-V4.1-Flash (deepseek-ai/DeepSeek-V4.1-Flash), trending #28 on Hugging Face as of 2026-09-10; reports span 2026-09-10 to 2026-09-12.
• Architecture: novel causal encoder-decoder design; CSA2 (Compressed Sparse Attention 2) shares KV across layers in Full, Reindex and Reuse modes; includes a 196B-parameter Engram conditional-memory module (Reports 1-2).
• Size (conflicting): 552B backbone plus 196B Engram parameters (Reports 1-2), 552B total (Report 3), or 763B total (Report 4). All sources agree: 8B active parameters at prefill, 16B at decode.
• Context window: 1M tokens, extended after sparse-attention training at 64K sequence length (Report 1).
• KV cache: 890 bytes per token — ~1/4 of DeepSeek-V4-Flash in fast GPU memory, one-eighth when offloaded (Reports 1-3; Report 4 cites 'up to 1/8'), and 437x smaller per token than DeepSeek-V1 (Reports 2-3), via FP4 (E2M1) quantization.
• Training: from scratch on 45T multimodal tokens (text and images); post-training via SFT, RL and on-policy distillation, with automated synthesis of agentic tasks and controllable reasoning effort (1-100).
• Benchmarks: 90.6 on Terminal-Bench 2.1 (Report 2); 74.2 on DeepSWE v1.1 (Reports 2-3), narrowly ahead of Opus 5 and GPT-5.6 Sol; Artificial Analysis Intelligence Index of 40, above DeepSeek V4 Pro 0813 (Report 4); ranked #1 open-weight on…
• Availability: MIT-licensed weights on Hugging Face; vLLM and SGLang support (Report 2); Baseten day-0 support and Ollama rollout to paid subscribers (Report 4); API at $0.30/$1.20 per 1M input/output tokens with 50% off-peak discount…