🤖 DeepSeek-V4.1-Flash is bigger than V4-Flash and needs about a quarter of the KV cache
DeepSeek's paper describes a multimodal MoE with a 1M-token context:
• 552B backbone params, using a Causal Encoder-Decoder design
• 8B active per token in prefill, 16B in decode
• Roughly 1/4 the runtime KV cache and 1/8 the persistent KV cache of V4-Flash
• 437x smaller per-token cache than DeepSeek-V1
It's cheap where agents hurt most, which is input-heavy prompts.
DeepSeek's paper describes a multimodal MoE with a 1M-token context:
• 552B backbone params, using a Causal Encoder-Decoder design
• 8B active per token in prefill, 16B in decode
• Roughly 1/4 the runtime KV cache and 1/8 the persistent KV cache of V4-Flash
• 437x smaller per-token cache than DeepSeek-V1
It's cheap where agents hurt most, which is input-heavy prompts.
arXiv.org
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill...
❤1
🚨 The DOJ is asking questions about Nvidia buying Hugging Face
Reuters reports that antitrust staff have started inquiring about the deal, citing anonymous sources. The biggest AI chip company would own the place where open models live.
It's an inquiry, not a formal challenge. No timeline has been reported for the deal.
Reuters reports that antitrust staff have started inquiring about the deal, citing anonymous sources. The biggest AI chip company would own the place where open models live.
It's an inquiry, not a formal challenge. No timeline has been reported for the deal.