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

It splits three ways in July 2026: Kimi K3 leads the Arena.ai Frontend Code Arena (ahead of Claude Fable 5), GLM-5.2 tops the open-weights Intelligence Index and is the most deployable (744B, MIT license), and DeepSeek V4 leads raw SWE-bench Verified and competitive coding. For price-to-performance, DeepSeek V4 Pro wins; for the highest ceiling, Kimi K3; for self-hosting, GLM-5.2.

Which open-weight model is cheapest for coding?

DeepSeek V4 Flash at $0.14 input / $0.28 output per MTok off-peak is the cheapest, with DeepSeek V4 Pro at $0.435/$0.87 off-peak for harder tasks (both 2x during peak hours 1-4 and 6-10 UTC). GLM-5.2 is $1.40/$4.40. Kimi K3 is the most expensive open model at $3/$15.

Which open-weight coding model is easiest to self-host?

GLM-5.2 (744B, MIT license) is the most practical to self-host of the frontier open models — self-hostable today on realistic multi-GPU hardware. DeepSeek V4 Pro (1.6T) and Kimi K3 (2.8T) need data-center clusters. For a single 24GB card, DeepSeek V4 Flash or a Qwen3.6 27B-class model after INT4 quantization are the realistic picks.

Quick Answer

Best Open-Weight Coding Model 2026: Ranked

Published: July 26, 2026

The Short Answer

In July 2026 the open-weight coding crown splits three ways:

Kimi K3 — highest ceiling. Leads the Arena.ai Frontend Code Arena, Artificial Analysis Index ~57. But $3/$15 API and 2.8T weights make it the priciest and hardest to self-host.
GLM-5.2 — best all-round. 744B, MIT license, tops the open-weights Intelligence Index, self-hostable today, $1.40/$4.40.
DeepSeek V4 — best value. Leads raw SWE-bench Verified and competitive coding; $0.435/$0.87 (Pro) or $0.14/$0.28 (Flash) off-peak.

Pick GLM-5.2 to self-host, DeepSeek V4 for cheapest API coding, Kimi K3 for the top quality ceiling.

The Ranking Table (July 2026)

Model	$/MTok in	$/MTok out	License	Self-host	Coding strength
DeepSeek V4 Pro	$0.435*	$0.87*	Open weights	Data-center	SWE-bench, competitive
DeepSeek V4 Flash	$0.14*	$0.28*	Open weights	24GB (quantized)	Fast, cheap tasks
GLM-5.2	$1.40	$4.40	MIT	Moderate	Agentic, long-context
Kimi K3	$3.00	$15.00	Modified MIT	Multi-node	Frontend Code Arena #1
Qwen 3.7 Max	varies	varies	Open weights	Data-center	Strong all-round

DeepSeek off-peak; 2x during peak hours (1-4 & 6-10 UTC). V4 Pro cache-hit ~$0.043.

Best by Use Case

Best value coding → DeepSeek V4 Pro / Flash. Nothing beats $0.435/$0.87 (Pro) for cost-per-task on real coding work, and V4 leads raw SWE-bench Verified plus LiveCodeBench (~93.5%) and Codeforces (~3206). Use Flash ($0.14/$0.28) for high-volume, cheap tasks and Pro for hard reasoning. Watch peak-hour surge and note the July 24 retirement of the deepseek-chat/deepseek-reasoner names.

Best to self-host → GLM-5.2. The 744B MoE with a true MIT license is the only frontier-class open model you can realistically run, fine-tune, and ship commercially without license friction. Tops the open-weights Intelligence Index and serves faster than K3 or V4 Pro. Output cost ($4.40) is ~1/3 of Kimi K3’s.

Best quality ceiling → Kimi K3. Frontend Code Arena leader (ahead of Claude Fable 5), Index ~57. Full weights drop July 27 under Modified MIT, but at 2.8T you need a ~16x H100 cluster to self-host and $3/$15 to use the API. Choose it when quality outranks cost.

Strong all-rounder → Qwen 3.7 Max. Alibaba’s flagship remains a capable open coding model and a common Claude Code drop-in, though it trails the top three on the most-cited July 2026 coding benchmarks.

How to Choose

Cheapest API coding → DeepSeek V4 Pro (or Flash for volume)
Self-host on your own hardware → GLM-5.2 (MIT)
Highest open-weight quality → Kimi K3
Commercial license clarity → GLM-5.2 (MIT) or DeepSeek V4
Single 24GB card → DeepSeek V4 Flash (quantized)

Sources

Best open-source coding model 2026 (Morph): morphllm.com/best-open-source-coding-model-2026
Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2 (MarkTechPost, July 18, 2026): marktechpost.com
GLM-5.2 vs DeepSeek V4 Pro (CodingFleet): codingfleet.com/blog/glm-5-2-vs-deepseek-v4-pro
DeepSeek V4 Pro pricing (OpenRouter): openrouter.ai/deepseek/deepseek-v4-pro