Which open-weight model is best to run locally in August 2026?

For a single high-end GPU or workstation, Qwen3.8-27B (open weights due the week of Aug 10, 2026) is the freshest option and pairs with Alibaba's strong agentic tuning. GLM-5.2 is a proven, well-quantized workhorse with permissive licensing. Kimi K3 (open weights July 27, 2026) is strongest for long-context coding but heavier. Pick by VRAM and task, not just benchmarks.

Do these run on a single consumer GPU?

Qwen3.8-27B is designed for single high-end GPUs (24-48GB) with quantization; GLM-5.2 has widely available 4-bit quants that fit 24GB cards. Kimi K3's full weights are larger and typically need more VRAM or multi-GPU, though quantized community builds narrow the gap. Always check the exact quant and context length you need.

Is Qwen3.8-Max the same as Qwen3.8-27B?

No. Qwen3.8-Max is the 2.4-trillion-parameter MoE flagship that needs a multi-GPU cluster — not a local model. Qwen3.8-27B is the smaller sibling built for single-GPU self-hosting, releasing open-weight the same week (week of Aug 10, 2026). For local use, you want the 27B, not the Max.

Quick Answer

Qwen3.8-27B vs GLM-5.2 vs Kimi K3 (Aug 2026)

Published: August 6, 2026

The Short Answer

For self-hosting on your own hardware in August 2026: GLM-5.2 is the proven, easy-to-quantize workhorse available today; Kimi K3 (open weights July 27, 2026) leads long-context coding but is heavier; Qwen3.8-27B (open weights due week of Aug 10, 2026) is the freshest single-GPU option with Alibaba’s agentic tuning. Choose by VRAM and task.

Side-by-Side

	Qwen3.8-27B	GLM-5.2	Kimi K3
Vendor	Alibaba	Zhipu	Moonshot
Open weights	Due week of Aug 10, 2026	Available	Available (Jul 27, 2026)
Local fit	Single high-end GPU (24-48GB)	24GB with 4-bit quant	Heavier; multi-GPU or aggressive quant
Best at	Agentic tasks, multimodal, fresh tuning	General coding, reliability, quant support	Long-context coding
License	TBD at launch	Permissive	Open weights

How To Choose

Best long-context coding you can host → Kimi K3 — strong on large codebases, but budget the VRAM (K3 API is $3/$15 if you’d rather not host).
Easiest, most reliable single-GPU workhorse → GLM-5.2 — mature quantization, broad tooling support, permissive license. The safe default.
Freshest, agentic, multimodal → Qwen3.8-27B — wait for the week-of-Aug-10 drop and the license terms, then test. Pairs conceptually with the Qwen3.8-Max flagship.

The Trade-Offs

Availability: GLM-5.2 and Kimi K3 are downloadable now; Qwen3.8-27B is imminent, not shipped, as of Aug 6, 2026.
License clarity: GLM-5.2 and Kimi K3 have known terms. Qwen3.8-27B’s license was undisclosed at the Aug 3 Max launch — verify before commercial use.
VRAM reality: “27B” and “K3” mean very different memory footprints. Match the quant to your card before downloading.

Verdict

Ship today, single GPU → GLM-5.2
Long-context coding → Kimi K3
Freshest agentic pick (pending drop + license) → Qwen3.8-27B

Sources

Medium — Qwen3.8-27B local release preview: medium.com
MarkTechPost — Qwen3.8 family release: marktechpost.com
Latent Space — Qwen3.8-Max 2.4T and 27B analysis: latent.space