Which open-weight AI model is best in August 2026?

It depends on your constraint. Muse Glimmer (Meta, 30B dense, Apache 2.0) is best for truly unencumbered local agents. Qwen3.8-Max (2.4T MoE) is the most capable — #1 in reasoning on some boards — but attaches revenue-share terms for large commercial users. Kimi K3 leads SWE Marathon but also carries revenue-share strings above $20M annual revenue.

Which open-weight model has the most permissive license?

Muse Glimmer. It ships under Apache 2.0 — a permissive, OSI-approved license with no revenue share. Qwen3.8-Max and Kimi K3 both attach commercial revenue-sharing terms (Kimi requires businesses over $20M annual revenue to negotiate, potentially up to a 30% share), so they are 'open-weight' rather than open source.

Can I run these on consumer hardware?

Muse Glimmer's 30B dense design targets consumer hardware and always-on local agents. Qwen3.8-Max (2.4T parameters) and Kimi K3 are far larger and need serious multi-GPU infrastructure — they are cloud/self-host-at-scale models, not laptop models.

Quick Answer

Muse Glimmer vs Qwen3.8-Max vs Kimi K3 (Aug 2026)

Published: August 12, 2026

The Short Answer

As of August 2026: Muse Glimmer (Meta, 30B dense, Apache 2.0) wins on license and local footprint. Qwen3.8-Max (2.4T MoE) is the most capable — #1 reasoning on some boards — but carries revenue-share strings. Kimi K3 leads SWE Marathon and also attaches commercial terms. Pick by freedom, capability, or coding.

Quick Comparison

	Muse Glimmer	Qwen3.8-Max	Kimi K3
Vendor	Meta	Alibaba	Moonshot AI
Size	30B dense	2.4T MoE	Very large MoE
License	Apache 2.0 (no strings)	Open-weight + revenue share	Modified MIT + revenue share
Strength	Local agents, footprint	Reasoning, agentic, multimodal	SWE Marathon #1
Runs on	Consumer hardware	Multi-GPU / cloud	Multi-GPU / cloud

Muse Glimmer — Freedom & Footprint

30B dense, multimodal, Apache 2.0 with no revenue share (released Aug 10, 2026). The pick when you need to ship and self-host without legal friction, or run always-on agents on consumer GPUs. Won’t top frontier boards, but nothing else here is this unencumbered and this small.

Qwen3.8-Max — Most Capable

2.4T MoE (released Aug 3, weights Aug 12, 2026), multimodal, strong on OSWorld-Verified (86.1) and ranked #1 reasoning on BenchLM’s board. But Alibaba plans revenue-sharing for large commercial users — so it’s “open-weight,” not OSI open source. The pick when you need maximum capability and can live with the terms.

Kimi K3 — Coding Leader

Kimi K3 (Moonshot) tops SWE Marathon and is among the largest open-weight models. Its modified MIT license requires businesses over $20M annual revenue to negotiate — up to a 30% revenue share. The pick for agentic coding at scale if you’re under the revenue threshold.

Which Should You Pick?

Truly permissive, local, small → Muse Glimmer.
Maximum capability, terms OK → Qwen3.8-Max.
Best open-weight coding → Kimi K3.

The Reality Check

The 2026 open-weight story is licenses, not just benchmarks. Two of these three attach revenue-share strings that make them non-OSI. If your business needs genuinely free commercial use, Muse Glimmer’s Apache 2.0 is the differentiator — even if it gives up raw capability.

Sources

AINews — Meta Muse Glimmer (30B, Apache 2.0): news.smol.ai
DataCamp — Qwen3.8-Max: datacamp.com
Tech in Asia — Alibaba plans revenue sharing for Qwen: techinasia.com