DeepSeek released V4-Pro (1.6T params, 49B active) and V4-Flash (284B/13B), both with 1M-token context and radically cheaper inference for long agentic workloads.
DeepSeek released two MoE checkpoints on Hugging Face: V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B total / 13B active), both supporting 1M-token context windows. V4-Pro runs at 27% of DeepSeek-V3.2's per-token FLOPs and uses 10% of its KV cache memory; V4-Flash drops to 10% FLOPs and 7% KV cache. Compared to standard grouped query attention in bfloat16, V4 requires approximately 2% of KV cache memory. The architecture introduces Compressed Sparse Attention (CSA) with 4x sequence compression and a new multi-mode reasoning system (Non-Think, Think High, Think Max) designed explicitly for long agentic task trajectories.
The CSA architecture fundamentally breaks the cost curve for long agentic traces. At 1M tokens, V4-Pro uses 27% of the FLOPs of V3.2 and 2% of the KV cache vs. standard GQA in bfloat16 — meaning you can run multi-step tool-use trajectories (SWE-bench, browser agents, terminal sessions) that previously OOM'd or required expensive chunking hacks. The three-mode reasoning system (Non-Think / Think High / Think Max) also gives you inference-time control over compute budget without separate model deployments.
Pull V4-Flash from Hugging Face and run your longest failing agent trace through it this week — measure whether the tool-call degradation at step 50+ disappears without any prompt engineering changes.
Install: pip install transformers accelerate and pull the model with from transformers import AutoModelForCausalLM, AutoTokenizer; tokenizer = AutoTokenizer.from_pretrained('deepseek-ai/DeepSeek-V4-Flash')
Tags
Related