Which open-weight model is best to self-host in July 2026?

For raw intelligence, Kimi K3 leads (Artificial Analysis Intelligence Index ~57, #3 overall behind only Claude Fable 5 and GPT-5.6 Sol). For a customizable, permissively-licensed base to fine-tune, Thinking Machines' Inkling (Apache 2.0, 975B total / 41B active) is the pick. For lowest serving cost, DeepSeek V4 Pro is near the floor.

Is Inkling trying to beat GPT-5 or Claude?

No. Thinking Machines explicitly says Inkling (released July 15, 2026) is not aimed at topping closed-model benchmarks. It's a robust, well-rounded 975B-parameter MoE base trained on 45T multimodal tokens, meant for developers and enterprises to download, run locally, and fine-tune under Apache 2.0.

Is Kimi K3 open weight and how much does the API cost?

Kimi K3 released open weights on July 27, 2026. Moonshot's hosted API is flat-priced at $3/$15 per MTok (no peak/valley pricing). Self-hosting is possible but K3 is trillion-scale MoE, so serving needs serious GPU memory.

Quick Answer

Inkling vs Kimi K3 vs DeepSeek V4 Pro to Self-Host

Published: July 29, 2026

The Short Answer

Pick by goal: Kimi K3 for the highest open intelligence, Inkling for a permissive, fine-tunable base, DeepSeek V4 Pro for lowest serving cost. All three are open-weight and released in July 2026.

The Comparison

Model	Released	Params	License	Intelligence Index	Hosted API
Kimi K3	Jul 27	Trillion-scale MoE	Open weights	~57 (#3 overall)	$3 / $15
Inkling	Jul 15	975B total / 41B active	Apache 2.0	well-rounded base	— (self-host)
DeepSeek V4 Pro	2026	Trillion-scale MoE	Open weights	~44 (Max)	~$0.435 / $0.87*

DeepSeek off-peak; 2x during peak (1–4 & 6–10 UTC). Cache-hit ~$0.043.

Kimi K3 — highest open intelligence

Moonshot’s K3 ranks #3 overall on the Artificial Analysis Intelligence Index (~57), behind only Claude Fable 5 and GPT-5.6 Sol, and comparable to Opus 4.8. It adds native vision. Demand at launch was so high Moonshot briefly suspended new subscriptions. Flat $3/$15 pricing, open weights July 27.

Inkling — the customizable base

Mira Murati’s Thinking Machines released Inkling on July 15, 2026: a 975B-total / 41B-active MoE, 1M context, trained on 45T multimodal tokens (text/image/audio/video). It’s explicitly not chasing top closed-model scores — it’s a well-rounded base for fine-tuning, licensed Apache 2.0, with an Inkling-Small (12B active) preview and Tinker console support.

DeepSeek V4 Pro — the cost floor

Text-only, trillion-scale MoE, priced near the absolute floor (~$0.435/$0.87 off-peak, 2x peak). Intelligence trails Kimi K3 but the serving economics are unmatched for high-volume text pipelines.

What to Do

Need the smartest open model: Kimi K3.
Building a fine-tuned product: Inkling (Apache 2.0 is the friendliest license here).
Cheapest high-volume text: DeepSeek V4 Pro.

Sources

Thinking Machines — Introducing Inkling: thinkingmachines.ai/news/introducing-inkling
MarkTechPost — Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: marktechpost.com
Artificial Analysis — Kimi K3 vs DeepSeek V4 Pro: artificialanalysis.ai/models/comparisons/kimi-k3-vs-deepseek-v4-pro