self-hosting - PacketNebula

Official Qwen architecture diagram for Qwen3.8-Flash-Next, showing input tokens feeding a vocabulary embedding and a separate n-gram embedding layer at layer 2, a stack of Gated DeltaNet layers punctuated by Qwen Sparse Attention layers in a three to one hybrid block, Gated Residual read and write gates around each MoE block, and MTP modules beside the prediction head.

Somebody on the team read "6B activated" and asked which of our boxes could run it. Fair question, wrong number: the FP8 checkpoint for Qwen3.8-Flash-Next is about 173 GiB of ...

Thinking Machines Lab announcement artwork for Inkling-Small, hand-drawn black loops winding across a cream graph-paper grid, carrying no text or product imagery.

Inkling-Small out-codes the 975B flagship at 276B

by stephane

3 September 2026

Two weeks ago we said the honest problem with Inkling was that you probably couldn't run it. Two terabytes of VRAM will do that. Thinking Machines seems to have agreed, ...

Poolside announcement graphic for Laguna S 2.1, purple isometric cubes descending across a grid on a cream background, with the words Introducing Laguna S 2.1 in black.

Laguna S 2.1: 118B open weights on one DGX Spark

by stephane

28 July 2026

Somebody forwards you the headline: an open model that beats rivals ten times its size. Then you open Poolside's own benchmark table and Kimi K3 is sitting there at 88.3 ...

Voxel Colosseum scene rendered in a browser demo generated by Kimi K3, showing a Roman amphitheatre packed with crowds inside a blocky city, with an FPS 120 overlay in the corner.

The Kimi K3 weights are 1.4 TB, and that is the catch

by stephane

3 September 2026

Somebody on your team is going to ask whether you can just host Kimi K3 yourselves. Short version: the weights land by July 27, and they weigh roughly 1.4 terabytes. ...

Answer card: Fireworks closed a 1.505 billion dollar Series D on July 16 2026 at a 17.5 billion dollar valuation, led by Atreides Management, Index Ventures and TCV with Nvidia among the backers, as an AI inference platform for serving and fine-tuning open-weight models.

Fireworks raised $1.5B, and a raise ships no code

by stephane

3 September 2026

You read our piece on Qwen 3.8, nodded at the open weights, then hit the same wall we always do. Great, now where do you actually run a near trillion-parameter ...

Answer card: Thinking Machines released Inkling on July 15 under Apache 2.0, a 975B mixture-of-experts model with 41B active parameters per token and a 1M token context window, which the lab says is not the strongest model available today.

Inkling is a 975B open model that loses benchmarks on purpose

by stephane

3 September 2026

Mira Murati's lab spent about nine months in near silence, and on July 15 it shipped weights instead of a waitlist. Inkling is a 975B mixture-of-experts model under Apache 2.0, ...

Answer card: Tencent open-sourced Hy3, an Apache 2.0 licensed 295B mixture-of-experts model that runs only 21B active parameters per token, with 256K context, free on OpenRouter until July 21.

Is Tencent’s 295B Hy3 really free?

by stephane

3 September 2026

Another giant model landed on Hugging Face, and this one is worth stopping for. On July 6 Tencent open-sourced Hy3, a 295B mixture-of-experts model under a plain Apache 2.0 license, ...