Which open-weight model is best for agentic coding in August 2026?

For proven, deployable open weights today, Kimi K3 (open weights July 27, 2026) is the safest agentic coding pick, with GLM-5.2 strong for long-horizon coding on more modest VRAM. Qwen3.8-Max (2.4T MoE, weights ~week of Aug 10) claims the top agentic benchmarks but is unverified and needs heavy GPUs. Deploy Kimi K3 or GLM-5.2 now; watch Qwen3.8-Max.

Is Qwen3.8-Max better than Kimi K3?

On Alibaba's claimed benchmarks, Qwen3.8-Max reports leads (OSWorld-Verified 86.1, Terminal-Bench 2.1 86.6) and ranked near Kimi K3 in early third-party arenas. But Kimi K3's weights have been public since July 27 and are battle-tested, while Qwen3.8-Max's weights hadn't shipped as of August 5. Kimi K3 is the proven choice; Qwen3.8-Max is the claimed leader.

Which of these is easiest to self-host?

GLM-5.2 is the most approachable for mid-VRAM setups, Kimi K3's weights are already available and well-supported, and Qwen3.8-Max is the hardest — a 2.4-trillion-parameter MoE needs large multi-GPU memory. Match the model to your hardware before chasing the top benchmark.

Quick Answer

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

Published: August 5, 2026

The Short Answer

For open-weight agentic coding in August 2026: Kimi K3 is the proven, deployable leader (weights since Jul 27); GLM-5.2 is the best mid-VRAM long-horizon coder; Qwen3.8-Max claims the top agentic benchmarks but is unverified and GPU-hungry (weights ~Aug 10). Deploy Kimi K3 / GLM-5.2 now, watch Qwen3.8-Max.

Side-by-Side

	Qwen3.8-Max	Kimi K3	GLM-5.2
Vendor	Alibaba	Moonshot	Zhipu
Open weights	~week of Aug 10	Yes (Jul 27)	Yes
Params	2.4T (MoE)	Large MoE	Mid MoE
Terminal-Bench 2.1*	86.6	—	—
SWE-bench Pro*	67.7	—	—
API price (in/out)	TBD	$3 / $15	Low
Self-host difficulty	Very high	Moderate	Approachable

*Qwen3.8-Max figures are Alibaba-reported, unverified independently as of Aug 5, 2026.

How To Choose

Ship an open-weight agent today → Kimi K3. Weights are public and mature; $3/$15 API if you don’t self-host. Best proven agentic coder in this group.
Self-host on modest hardware → GLM-5.2. Strong long-horizon coding without data-center GPUs.
Chase top claimed benchmarks → Qwen3.8-Max, once weights ship and independent SWE-bench/OSWorld numbers confirm Alibaba’s claims — and if you have the GPUs for 2.4T.

The Trade-Off

Proven vs claimed: Kimi K3 and GLM-5.2 are deployed and tested; Qwen3.8-Max’s lead is vendor-reported.
Capability vs runnability: Qwen3.8-Max aims highest but is the hardest to serve; GLM-5.2 is the pragmatic pick.
API vs self-host: Kimi K3’s $3/$15 API is a fine hedge until you commit GPUs.

Watch Outs

Qwen3.8-Max weights + license unconfirmed at launch (verify before planning self-host).
Benchmark tables lagged the Qwen3.8-Max launch — some are internal evals.

Verdict

Best proven open agentic coder → Kimi K3
Best mid-VRAM self-host → GLM-5.2
Top claimed / watch → Qwen3.8-Max

Sources

MarkTechPost — Qwen3.8-Max: marktechpost.com
Moonshot — Kimi: kimi.moonshot.cn
Zhipu AI — GLM: z.ai