DeepSeek V4.1 Flash lands on AA-AgentPerf with NVIDIA and AMD results

AA-AgentPerf replays real agentic coding sessions, including prompt caching and tool-call delays, and reports throughput per megawatt. This gives concrete hardware and cost data for serving open-weights models to agents.

• Artificial Analysis says DeepSeek V4.1 Flash scores 39 on its Intelligence Index, above V4 Pro 0813 at 36, with about a third of the parameters: a 552B backbone and 16B active versus 1.6T and 49B.
• The architecture is a causal encoder-decoder that activates 8B parameters per input token and 16B per output token, plus 196B of Engram n-gram memory that can live in host memory instead of on the accelerator.
• Compressed sparse attention needs roughly a quarter of the earlier V4 Flash KV cache per token, which Artificial Analysis says leaves room for more concurrent agents per GPU.
• Because the model runs on as few as 2 GPUs, an 8-GPU B300 node stays within about 20% of rack-scale GB300 NVL72; at 100 tok/s per user GB300 leads at ~261M tokens/s per MW versus ~128M for AMD MI355X.
• AA-AgentPerf replays real agentic coding sessions, including prompt caching and tool-call delays, and reports throughput per megawatt alongside per-agent speed to show the utilization versus user-experience trade-off.

• attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
• prompt caching — Reusing the model's processed form of a repeated prompt prefix so subsequent calls skip re-reading it, cutting cost and latency substantially.
• KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
• AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

• postSonnet 5.5 effort levels: max tops Coding Agent Index, xhigh saves 77%
• postCache hit rates explain a 9x provider cost gap for agentic tasks
• postFish Audio Drama 3 Preview leads open-weights TTS
• postGemini 4 Argon GA: frontier parity at lower cost per task

• postDeepSeek V4.1 Flash Beats Its Own Flagship on Price and PerformanceArtificialAnlys
• postDeepSeek V4 Pro 0813 scores 53 on the Artificial Analysis Intelligence Index, 8…Artificial Analysis
• postDeepSeek Ships V4.1-Flash With Native Vision, Smaller KV Cachedeepseek_ai

Checking sign-in…

Loading comments…