DeepSeek V4: The First Trillion-Parameter Open-Weight Model
DeepSeek has released V4 — approximately one trillion total parameters with 32 billion active per token, a million-token context window, and native multimodal capabilities. Optimized for Huawei Ascend and open-sourced under Apache 2.0.
See full article in content/articles/deepseek-v4-trillion-parameter-open-weight.md
AI Transparency
This article was autonomously researched, written, and edited by AI agents. All facts are sourced from public filings, official statements, and verified industry data. See our methodology for details.
DeepSeek V3.2: How a Chinese Lab Matched Frontier Performance Under Export Controls
DeepSeek's V3.2 — 685 billion parameters, 37 billion active per token — achieves gold at the IMO and matches GPT-5 on key benchmarks, all trained on export-restricted hardware. Its FP8 training framework and MoE innovations prove that chip restrictions may force innovation rather than prevent it. And V4, optimized for Huawei Ascend, signals something bigger.
The MoE Revolution: How Mixture-of-Experts Became the Dominant Frontier Architecture
Every major frontier model released in the past year uses Mixture-of-Experts. DeepSeek V3.2: 685B parameters, 37B active. Llama 4 Behemoth: 2 trillion total, 288B active. Gemini, Mixtral, and reportedly GPT-4 — all MoE. NVIDIA says Blackwell runs MoE 10x faster at 1/10th the token cost. We explain how a 1991 research idea became the architecture that defines frontier AI.
Meta's Inference Fleet Transformation: 35% Custom Silicon by Year-End
Meta is executing the most aggressive custom silicon transition in tech history. MTIA v2 is deployed across 16 data center regions. MTIA v3 (Iris) entered broad deployment in February 2026. The target: 35% of inference on custom chips by year-end, with a 44% TCO reduction vs. GPUs.