Which is the best 30B open-weight model for local agents in August 2026?

Qwen3.8-27B leads on published agentic coding scores (DeepSWE 42.2, Terminal Bench 73.0, OSWorld 84.3) and adds native vision input. Nemotron 3.5 Lightning is the fastest, with 3B active parameters claiming up to 4x faster output. Muse Glimmer 30B is the simplest permissive dense option. All three are single-GPU models.

Which of these open-weight models has the best license?

All three are commercially usable. Qwen3.8-27B and Muse Glimmer ship under Apache 2.0, a permissive OSI-approved license with no revenue-share strings. Nemotron 3.5 Lightning uses OpenMDW-1.1, which also allows commercial use and additionally publishes training data and recipes — the most transparent provenance of the three.

Can I run these models on one consumer GPU?

Yes, all three are designed for it. Nemotron 3.5 Lightning activates only 3B of its 30B parameters per token so it is the lightest at inference; Qwen3.8-27B and Muse Glimmer are dense models around 27-30B that fit on a 24GB card with quantization.

Quick Answer

Qwen3.8-27B vs Muse Glimmer vs Nemotron 3.5 Lightning

Published: August 16, 2026

The Short Answer

Between August 10 and 14, 2026, three vendors shipped ~30B open-weight agent models within five days: Meta’s Muse Glimmer 30B, Nvidia’s Nemotron 3.5 Lightning, and Alibaba’s Qwen3.8-27B. All run on a single consumer GPU. They differ on the axis that actually matters for your build: Qwen wins capability, Nemotron wins speed, Muse Glimmer wins simplicity.

Quick Comparison

	Qwen3.8-27B	Muse Glimmer 30B	Nemotron 3.5 Lightning
Vendor	Alibaba	Meta	Nvidia
Released	Aug 14, 2026	Aug 10, 2026	Aug 11, 2026
Architecture	27.8B dense	30B dense	30B MoE + Mamba-2, ~3B active
License	Apache 2.0	Apache 2.0	OpenMDW-1.1
Openness	Weights	Weights	Weights + training data + recipes
Context	262K tokens	—	—
Multimodal	Yes (image + video in)	Yes	—
Hardware	Single 24GB GPU	Consumer GPU	Single consumer GPU
Headline scores	DeepSWE 42.2 · Terminal Bench 73.0 · OSWorld 84.3	—	Claimed 4× faster output, 30% faster task completion
Companion tooling	Hugging Face, ModelScope	Hugging Face	NeMo Switchyard router

Where Each Wins

Qwen3.8-27B — best raw capability

Alibaba’s numbers are the strongest published in the class: DeepSWE 13.3 → 42.2 and QwenSWEBench 49.3 → 79.0 over the previous generation, plus Terminal Bench 73.0 and OSWorld 84.3. The 262K context and native image/video input make it the only one of the three genuinely equipped for UI-driving agents that read screenshots. If your local agent needs to see and then act, this is the pick.

Caveat: these are vendor-reported evals, several judged by other frontier models rather than deterministic tests. Verify on your own tasks.

Nemotron 3.5 Lightning — best speed and provenance

Nvidia’s hybrid interleaves Mamba-2 state-space layers with MoE, activating only ~3B of 30B parameters per token. Nvidia claims up to 4× faster output and 30% faster task completion than comparable models. For agent loops — long context read repeatedly, many small steps, latency compounding — that architecture is aimed precisely at the bottleneck.

It also has the cleanest provenance: OpenMDW-1.1 plus published training data and recipes. For enterprises with license-review processes, that is worth more than a few benchmark points. Nvidia shipped NeMo Switchyard, an open-source router, alongside it — the intended pattern is Lightning handling cheap frequent steps locally and escalating hard ones to a frontier API.

Muse Glimmer 30B — best simplicity

Meta’s 30B dense, multimodal, Apache 2.0 model is the least exotic of the three, and that’s the point. No MoE routing quirks, no mandatory thinking tokens, no revenue-share terms. It’s Meta’s clearest re-entry into open weights since the Llama-to-Muse pivot, and a genuinely unencumbered base for startups shipping products on top.

It doesn’t lead any benchmark table here. It’s the boring, dependable option — which is often the right one.

The Decision Framework

Agent must read screens or images → Qwen3.8-27B. Only it pairs strong OSWorld performance with native vision input and 262K context.
Latency is the bottleneck (interactive loops, many small steps) → Nemotron 3.5 Lightning. The 3B-active design is a step change, not a marginal gain.
Legal review is the bottleneck → Nemotron 3.5 Lightning (data + recipes published) or Muse Glimmer (plain Apache 2.0).
You want one dense model to fine-tune and forget → Muse Glimmer 30B.
You want maximum capability regardless of hosting → none of these; use an API model like GLM-5.3 or Gemini 3.7 Flash.

The Honest Reality Check

All three of these are ~30B models, and ~30B models are not frontier models. They win when data residency, latency, or per-token cost at extreme volume dominates your decision. If none of those apply, a $0.75/$3.75 API model will beat all three on quality per unit of engineering effort — and you won’t be paying for idle GPU time.

The real 2026 pattern is both: a local ~30B handling the cheap 80% of steps, with a router escalating the hard 20% to a frontier API. That’s exactly what Nvidia built Switchyard for, and it’s the architecture worth designing toward. See how to cut AI API costs in 2026.

Last verified: August 16, 2026.