Xanu

Putting applied AI to work — backed by the data, infrastructure, and manufacturing to innovate end to end.

[Xanu @xanuinc]

Open-sourcing our DeepSeek-V4-Flash + DSpark setup for 2× DGX Spark (GB10).

Stock container, no custom image. The SM120 topk fix in the wild rounds 256 down to 128 and discards half the draft candidates — we pad up to 512 instead and get +32% decode.

Recipes + scripts:
https://t.co/0w5hNbOTYv

[Xanu @xanuinc]

Full frontier LLM — GLM-5.2, unpruned, all 256 experts — running on 4 small NVIDIA DGX Sparks. Locally on a shelf, not in a datacenter. 🖥️

📏 327K context · ⚡ ~25 tok/s · ☁️ multiple options

Open source on github - clone it, edit one line, run it 👇
https://t.co/KwRd6fMacm

[Xanu @xanuinc]

@cohere See the launch post for more details: https://t.co/ZjtazXdtP1

[Xanu @xanuinc]

Check out how XanuNetworks/North-Mini-Code-1.0-NVFP4 achieved 60.13 tokens/sec on text generation on NVIDIA DGX Spark with vLLM! Shout out to @cohere for making this great new agentic coding model, which now fits in only 17gb NVFP4.

View full benchmark at https://t.co/TmVjnqpQFL

[Xanu @xanuinc]

Check out our blog to see our latest updates: https://t.co/JfAwcxRDP4

[Xanu @xanuinc]

@xanuinc Stay tuned to our twitter account for updates from Xanu