What is the best open-weight AI model to self-host in 2026?

For raw intelligence and multimodal input, Kimi K3 (2.8T MoE, weights released July 27, 2026) leads at ~57 on the Artificial Analysis Intelligence Index. For coding and fast agentic workflows, GLM-5.2 tops SWE-bench Pro at 62.1% and runs at ~168 tokens/sec. For the cheapest, most accessible self-host, DeepSeek V4 Flash fits on a single 80GB GPU.

Which open-weight model is easiest to run on limited hardware?

DeepSeek V4 Flash (284B total, ~13B active) runs on a single 80GB GPU when quantized, making it the easiest frontier-class open model to self-host. Kimi K3 needs 64+ accelerators for production, and GLM-5.2 needs over 1TB of VRAM in BF16.

Are these open-weight models really free to use?

The weights are released under permissive licenses (MIT for DeepSeek V4 and GLM-5.2, modified MIT for Kimi K3), so you can run them commercially. But you pay for the GPUs and electricity to serve them — self-hosting frontier MoE models is expensive at scale even when the weights cost nothing.

Quick Answer

Best Open-Weight AI Model to Self-Host (July 2026)

Published: July 28, 2026

The Short Answer

There is no single winner — the best open-weight model to self-host in July 2026 depends on your constraint:

Best intelligence + vision → Kimi K3 (2.8T MoE, weights out July 27, 2026)
Best coding + speed → GLM-5.2 (62.1% SWE-bench Pro, ~168 tok/s)
Cheapest + easiest to run → DeepSeek V4 Flash (single 80GB GPU)

The Three Contenders

	Kimi K3	GLM-5.2	DeepSeek V4
Maker	Moonshot AI	Z.ai	DeepSeek
Weights released	Jul 27, 2026	Jun 17, 2026	mid-Jul 2026
Params	~2.8T MoE	744B MoE (~40B active)	Pro 1.6T / Flash 284B
Context	1M	1M	1M
Vision	✅	❌	❌
Standout	AA Index ~57 (#1 open)	SWE-bench Pro 62.1%	cheapest API, easy Flash deploy
License	modified MIT	MIT	MIT

Pick by Use Case

Frontier general intelligence, multimodal apps → Kimi K3. It edges every other open model on the Artificial Analysis Intelligence Index and is the only one of the three with native image/video input. The catch is deployment scale — plan for a supernode (64+ H100/B200).
Coding agents, long-horizon dev work, fast inference → GLM-5.2. It beats DeepSeek V4 Pro (55.4%) and GPT-5.5 (58.6%) on SWE-bench Pro, hits 82.7% on Terminal-Bench 2.1, and is the fastest of the three. Text-only, and it burns more output tokens per task.
Cost-sensitive, limited GPUs, self-host today → DeepSeek V4 Flash. One 80GB GPU when quantized, MIT license, and the API fallback ($0.14/$0.28 per MTok off-peak) is the cheapest frontier-class option if you’d rather not host at all.

Don’t Forget the Hosting Bill

Open weights are free; serving them is not. Kimi K3 at production scale is a multi-server, multi-GPU commitment. GLM-5.2 needs 1TB+ VRAM in BF16. Only DeepSeek V4 Flash is realistic on a single high-end card. For most teams, an OpenRouter/Together/Fireworks-style hosted endpoint is cheaper than owning the hardware until you hit serious volume.

Sources

Moonshot AI — Kimi K3: github.com/MoonshotAI/Kimi-K3
Z.ai — GLM-5.2: z.ai
DeepSeek — V4 model docs: deepseek.ai/deepseek-v4
MarkTechPost — open trillion-scale MoE comparison (July 18, 2026): marktechpost.com