2026-08-03

1. Two Western open-weight releases had been sitting unrecorded for three weeks

An Interconnects recap on 2026-08-02 named three models as the current open Pareto frontier. One of them, Kimi K3, this wiki has held since July. The other two it had never heard of.

Inkling — Thinking Machines Lab, 2026-07-15. A 975B total / 41B active multimodal MoE, 1M context, Apache 2.0, full weights on Hugging Face in BF16 and NVFP4, with day-zero support in SGLang, vLLM, llama.cpp and Unsloth. Pretrained on 45 trillion tokens; the MoE design largely follows DeepSeek-V3. The lab explicitly declines to call it the strongest model — its stated role is a customizable base for fine-tuning, sold alongside the lab's own Tinker platform.

Laguna S 2.1 — Poolside, 2026-07-21. 118B total / 8B active, built for agentic coding, up to 1M context, under OpenMDW-1.1. Poolside's own figures: 70.2% on Terminal-Bench 2.1 with thinking, 78.5% on SWE-Bench Multilingual, and 40.4% on DeepSWE v1.1 against DeepSeek-V4-Pro-Max's 9.0% at roughly one-sixth the active parameters.

The sharpest detail is a coincidence nobody drew. Thinking Machines Lab is a founding partner of the Open Secure AI Alliance, whose July 27 launch this wiki recorded alongside the standing objection that none of its members ships a frontier-class open model. Inkling's weights landed twelve days earlier and neither the announcement nor the coverage mentioned them — and on the one third-party measure held here it does not close the gap anyway: 41 on the Artificial Analysis Intelligence Index against Claude Opus 5 (max) at 61.

Why it matters: "restrict open weights" and "restrict Chinese models" had converged into nearly the same policy because the largest open releases were Chinese. Two US-addressed releases inside a week — one Apache 2.0, one pitched in coverage as "the West's answer to DeepSeek and Qwen" — put a domestic constituency on the side a distillation-focused enforcement regime would burden.

→ Inkling (new), Laguna S 2.1 (new), Thinking Machines Lab (new), Poolside (new), Open-Weights Policy Fight → Inkling snapshot · Laguna snapshot · Interconnects #23 · AA leaderboard 2026-08-02

2. MiniMax shipped a video model and promised the weights — the weights did not come

MiniMax H3 launched 2026-07-31 (aliased Hailuo 3.0): a single transformer reading text, images, video and audio in one context and returning video with native stereo audio — 1440p, 4–15 seconds at whole-second granularity, 24 fps, where the previous generation offered a choice between 6 and 10. MiniMax states it set aside the Hailuo-02 architecture; the new H3-Omni Transformer splits understanding from generation during training because multimodal context tripled sequence-length variance, lifting end-to-end throughput by nearly 30%. Live in the API as MiniMax-H3 and in the Hailuo app at $0.13 per generated second at 2K.

It is marketed as open-weight with weights promised "in the coming days". A day after launch none had shipped.

Why it matters: this is the third model this wiki has recorded whose open-weight status is an announcement rather than a release — Qwen 3.8 Max (Preview) made the same promise on July 19 and has still not kept it. When a policy fight is being argued over what open weights do to the threat surface, a marketing claim and a downloadable checkpoint have stopped being the same thing, and the spec table now says unknown for the license rather than a license name.

→ MiniMax H3 (new), MiniMax, Open-Weights Policy Fight → Source · MiniMax · MarkTechPost