Which small open-weight agent model is best in August 2026?

It depends on the job. Meta's Muse Glimmer 30B leads agent orchestration benchmarks (75.5 on MCP Atlas). Nvidia's Nemotron 3.5 Lightning (released August 11, 2026) is the speed and efficiency pick — 30B total but only 3B active parameters, with claimed 4x faster output. Qwen3.6-27B is the safest for factual work with a much lower hallucination rate (49% vs Glimmer's 82% on AA-Omniscience).

Can these models run on a single consumer GPU?

Yes, all three target single-GPU deployment. Muse Glimmer 30B runs on a 24-32GB card with weights quantizing under 20GB. Nemotron 3.5 Lightning's 3B-active MoE design makes it the lightest at inference time. Qwen3.6-27B also fits a single high-VRAM consumer GPU.

Which has the most permissive license?

Nemotron 3.5 Lightning is the most open package: OpenMDW-1.1 license with commercial use allowed, plus Nvidia published the training data and training recipes — something neither Meta nor Alibaba does for these models.

Quick Answer

Nemotron 3.5 Lightning vs Muse Glimmer vs Qwen3.6-27B

Published: August 15, 2026

The Short Answer

August 2026 turned the ~30B local agent model class into a three-way race. As of August 15, 2026: Muse Glimmer 30B (Meta, Aug 10) is the agent-benchmark leader, Nemotron 3.5 Lightning (Nvidia, Aug 11) is the fastest and most openly licensed, and Qwen3.6-27B (Alibaba) is the lowest-hallucination factual pick. All run on one consumer GPU.

Quick Comparison

	Nemotron 3.5 Lightning	Muse Glimmer 30B	Qwen3.6-27B
Vendor	Nvidia	Meta	Alibaba
Released	Aug 11, 2026	Aug 10, 2026	2026 (current Qwen small line)
Size	30B total / ~3B active (MoE + Mamba-2)	30B	27B
MCP Atlas (agents)	— (too new for third-party scores)	75.5	62.5
AIME 2026 (math)	—	94.7	94.1
Hallucination (AA-Omniscience)	—	82%	49%
Claimed speed	Up to 4× faster output	~233 tok/s (RTX 5090, DFlash)	Standard
License	OpenMDW-1.1 + training data + recipes	Open weights	Open weights
Hardware	Single consumer GPU	24-32GB GPU	Single GPU

Nemotron 3.5 Lightning — Speed and Openness

Nvidia’s entry (Aug 11, 2026) is architecturally the odd one out: Mamba-2 state-space layers interleaved with a 3B-active MoE, tuned for agent loops where latency compounds. Nvidia claims 4× faster output and 30% faster task completion vs. similar models. It’s also the most open: OpenMDW-1.1, commercial use allowed, training data and recipes published, available on Hugging Face, Ollama, LM Studio, and Cline. Third-party agent benchmarks are still landing — treat vendor speed claims as claims. Full breakdown: What is Nemotron 3.5 Lightning?

Muse Glimmer 30B — Agent-Benchmark Leader

Meta’s Muse Glimmer 30B (Aug 10, 2026) posts the best measured agent scores in class: 75.5 MCP Atlas, 94.7 AIME 2026, strong DeepSearch QA — built for always-on local agent workflows with a 120K+ context. Its known weakness is hallucination (82% on AA-Omniscience), so keep it tool-grounded. See Muse Glimmer vs Qwen3.6-27B vs Gemma 4.

Qwen3.6-27B — The Factual Safe Pick

Qwen3.6-27B trails Glimmer on orchestration (62.5 MCP Atlas) but nearly matches it on math (94.1 AIME) and hallucinates far less (49%). When your agent writes answers rather than calling tools, that reliability gap matters more than orchestration scores.

Which Should You Pick?

Tool-heavy agent loops, latency-sensitive: Nemotron 3.5 Lightning (verify speed claims on your workload).
Best measured orchestration: Muse Glimmer 30B — with retrieval grounding.
Factual/knowledge tasks: Qwen3.6-27B.
Enterprise license diligence: Nemotron — data + recipes provenance is the cleanest.

The bigger story: the US labs (Meta, Nvidia) shipped both models in the same week to counter China’s open-weight lead — good news for anyone deploying local agents, since this class now improves monthly.

Last verified: August 15, 2026.