What is the best open-weight coding model in August 2026?

GLM-5.2 and Kimi K3 lead on quality — GLM-5.2 scored 87 and Kimi K2.7 Code 86 on a local coding benchmark, and Kimi K3 (2.8T params, 1M context) posts scores next to Claude Opus 4.8. DeepSeek V4 Flash 0731 is the cheapest hosted API at $0.14/$0.28 per MTok. Pick GLM-5.2 for the best self-host value, Kimi K3 for peak open quality, DeepSeek for cheapest API.

How big are these open-weight models?

Kimi K3 is the largest open-weight model ever at 2.8T parameters with a 1M context and always-on reasoning. GLM-5.2 is 744B total (40B active via MoE) with a 1M context. DeepSeek V4 Flash is 284B total (13B active). GLM-5.2's smaller active footprint makes it the most self-hostable of the three.

Can I self-host Kimi K3 and GLM-5.2?

Yes — both ship downloadable weights. GLM-5.2 (744B, 40B active) is the more practical self-host on limited GPU budgets. Kimi K3 at 2.8T params needs serious infrastructure but delivers frontier-adjacent quality. DeepSeek V4 also releases weights, though V4 Flash is most used via its ultra-cheap hosted API.

Quick Answer

Kimi K3 vs GLM-5.2 vs DeepSeek V4: Open-Weight Coding Aug 2026

Published: August 2, 2026

The Short Answer

For open-weight coding in August 2026: GLM-5.2 is the best self-host value (744B, 40B active), Kimi K3 is peak open quality (2.8T params, scores next to Opus 4.8), and DeepSeek V4 Flash 0731 is the cheapest hosted API at $0.14/$0.28 per MTok.

The Comparison

	Kimi K3	GLM-5.2	DeepSeek V4 Flash 0731
Params	2.8T	744B (40B active)	284B (13B active)
Context	1M	1M	1M
Weights	Open	Open	Released
Local coding bench	~86 (K2.7 Code)	87	Competitive
Best at	Peak open quality	Self-host value	Cheapest hosted API

Where Each Wins

Kimi K3 → peak open quality. The largest open-weight model ever (2.8T params), always-on reasoning, native multimodal, benchmark scores next to Claude Opus 4.8. Best for long agent loops that need sustained reasoning — but needs heavy infra to self-host.
GLM-5.2 → best self-host value. 744B with only 40B active via MoE means it fits realistic GPU budgets, and it scored 87 on a local coding benchmark (edging Kimi K2.7 Code’s 86). The pragmatic on-prem pick.
DeepSeek V4 Flash 0731 → cheapest hosted API. $0.14/$0.28 per MTok, 1M context, 13B active. If you don’t need to run weights yourself, nothing capable is cheaper.

Self-Host vs Hosted

On-prem, limited GPUs → GLM-5.2 (40B active is the sweet spot)
On-prem, max quality, big infra → Kimi K3
Cheapest bill, no ops → DeepSeek V4 Flash 0731 hosted API

Self-hosting trades per-token fees for GPU cost and ops overhead — worth it at high volume or under data-residency rules.

Verdict

Best self-host value → GLM-5.2
Peak open-weight quality → Kimi K3
Cheapest capable API → DeepSeek V4 Flash 0731

Sources

Moonshot AI — Kimi: moonshot.ai
Zhipu AI — GLM: z.ai
DeepSeek API pricing: api-docs.deepseek.com/quick_start/pricing