DeepSeek-V4-Pro ships with 1.6T-param MoE; open-source eval harness steals the show
DeepLearningAI · x · 2026-09-03
DeepSeek released DeepSeek-V4-Pro-0813: a 1.6T-param MoE (49B active) with 1M input / 384K output context, adjustable reasoning, tool calls, and MIT-licensed weights. It scores 53 on AA's Intelligence Index (3rd among open weights) but only slightly beats its Flash model. API prices rose: $1.32/$0.044/$3.96 per 1M input/cached/output tokens at peak, half off-peak. The bigger story may be its free open-source agent harness, which logs every tool call, system prompt, and subagent schedule so devs can reproduce benchmarks and fork their own harnesses.
More from Infra
- BofA reiterates Nvidia top pick: FY28 70% growth is a floor, stock at ~16x CY27 PE, lowest in a decade — firstadopter · 2026-09-03
- DGX Spark first impressions: latest CUDA install is needlessly hard — QuixiAI · 2026-09-03
- Datacenters forecast to double in 2027 and again in 2028 as economics reorient — abhiadesai · 2026-09-03
- Broadcom CEO: Anthropic and OpenAI to become its No.1 and No.2 custom AI chip clients — dinabass · 2026-09-03
- Hock Tan says power availability dictates the exact timing of compute capacity deployment — BenBajarin · 2026-09-03
- Perplexity open-sources its Mac inference server optimized for Qwen 3.6 on Apple Silicon — Specter_Origin · 2026-09-03