Tag: open-source

DeepSeek V4 ships two open-weight MoE models — a 1.6T Pro and a 284B Flash — with novel sparse attention, FP4 quantisation, 1M token context, and validated Huawei Ascend NPU support. Here's what actually changed.

Alibaba released Qwen3.6-35B-A3B on 16 April 2026, the first open-weight model in the Qwen3.6 series. The benchmarks show real gains in agentic coding, but the architecture is unchanged from Qwen3.5 and the red flags warrant scrutiny.

A 26-person startup spent $20M training a 400B MoE model on 2,048 B300 GPUs — and produced the strongest open reasoning model outside China. Trinity-Large-Thinking ranks #1 on τ²-Airline at 1/28th the cost of Claude Opus 4.6.

The Linux kernel now has official AI coding guidelines — an Assisted-by tag, a ban on AI Signed-off-by, and Sashiko for automated review. What changed, and what it means for open source.

Gemma 4 brings frontier-level multimodal intelligence to open-source — with models ranging from 2B to 31B parameters, MoE efficiency, and native audio support for edge devices.

How Peter Steinberger's personal AI assistant went from 500 stars to 358,000, what the architecture looks like, and why OpenAI hired its creator.