Nanbeige4.3-33B just dropped Open 33B dense model, Apache 2.0, built for agentic coding and cowork tasks. Vendor scores beat Qwen3.8-27B on most agent benchmarks. But no paper, no replication. Test it yourself. Link in the first reply below.

NVIDIA RTX 3000 cards may have had half their BF16 math speed locked A new Linux module, gpu-unlock, reports 25–80% faster BF16 GEMM, no driver patching. Real end-to-end gains look smaller, and it's unsupported. Test before you trust it. Link in the first reply below.

Spark-X2.6-4B just dropped 4B open model (Apache 2.0) built for on-device agents. Card claims it beats some 9B models on tool use, browsing and SWE tasks. Strong numbers, but mostly vendor-reported. Verify before you trust. Link in the first reply below.

Recursive just open-sourced Flint 39.8B MoE coding model, only ~2.5B active per token. Apache 2.0. Beats sparse MoE peers on agentic benchmarks (vendor-reported). Dense Qwen 3.8 still leads on some. Local coding is getting real. Link in the first reply below.

I interviewed Adel El Hallak, VP of NVIDIA Agent Toolkit, at GTC Berlin. What an agent is made of. How the Agent Toolkit works with LangChain and OpenClaw. What a blueprint packs in. The Agent Knowledge Foundation, and Kumo. The full interview will be out soon.

One petaflop, on the table. With Jensen Huang at GTC Berlin. The machines behind us are RTX Spark. In 2016 he delivered 1 petaflop and 128GB to OpenAI, and it took two people to carry. This one is a Windows PC.

Held Hugging Face’s Microduck after the keynote. An agent on RTX Spark set the course. Isaac Sim ran the digital twin. Training ran in the cloud. The policy came back onto the robot. Thank you, NVIDIA, for hosting me.

Pregame is on at the Tempodrom. Jensen Huang's GTC Berlin keynote is 11:00–13:00 CEST. Free online. Pregame: Sovereign and Enterprise AI, AI Native, AI Safety and Security, AI Factories, and Physical AI and Industrial AI. nvda.ws/4gaEPYe @NVIDIA @NVIDIAGTC @NVIDIAAI

GTC Berlin, four jobs, not one product. A Vera Rubin factory in a container can go from six months to 14 days on a prepared site, NVIDIA says. OpenShell keeps the agent off files it was not given. Post-train Nemotron on your own data. Humanoids, cars, radiotherapy and networks.

DeepSeek just open-sourced DeepInfer The inference engine it says powers its own serving. Apache 2.0, built for V4.1-Flash on NVIDIA SM100 and Huawei Ascend. But no KV cache system, no P/D disaggregation, no load balancer. Great baseline, not tur Link in the first reply below.