The short answer
● Inkling 2 Small — 594B total, 17B active, weights not yet released
● Kimi K3.1 Pro — 2.8T total, 104B active, weights on Hugging Face
● Cost per task — $1.18 for Inkling, per Artificial Analysis
● Coding — Kimi narrowly leads SWEBench Pro v2 Hard
● Available now — Kimi. Inkling runs only on Tinker
Both are being described as the leading open-weight model. Only one of them has published weights, which makes the comparison simpler than the coverage suggests.
|
Inkling 2 Small |
Kimi K3.1 Pro |
| Lab |
Thinking Machines |
Moonshot AI |
| Released |
6 November 2026 |
23 October 2026 |
| Parameters |
594B total, 17B active |
2.8T total, 104B active |
| Weights available |
No, Apache 2.0 promised |
Yes, modified MIT |
| Where to run it |
Tinker platform only |
Hosted or self-hosted |
| Input |
Text and image |
Multimodal |
| SWEBench Pro v2 Hard |
86.3% |
Narrowly ahead |
| Hosted price |
$1.18 per task |
$2.00 / $10.00 per million |
If you need weights today
Kimi K3.1 Pro, and it is not close. The weights are published at moonshotai/Kimi-K3.1-Pro under a modified MIT licence, so you can self-host. Modified means standard MIT semantics do not carry over, so read the file before you assume anything.
Inkling 2 Small is a proprietary offering on Tinker today. Thinking Machines says Apache 2.0 weights follow once external safety testing finishes, and has given no date. If your plan depends on self-hosting, you cannot plan around it yet.
If you are buying hosted inference
Artificial Analysis puts Inkling 2 Small at roughly 150,000 output tokens per task and $1.18 per task, against $7.77 for Gemini 3.8 Flash and $10.02 for Claude Sonnet 5.5. That is the strongest argument for it, and it holds whether the weights ever ship or not.
One figure to read carefully: Inkling’s hallucination rate fell from 61% to 27%, but its response rate fell from 77% to 53%. Part of the accuracy gain is the model answering fewer questions. Whether that is a feature depends entirely on your workload.
The honest verdict
Kimi K3.1 Pro if you want to run it yourself or you are optimising for coding. Inkling 2 Small if you are buying hosted inference and cost per task is the number that decides it. Neither answer changes until Thinking Machines actually publishes the weights.
Sources: The Decoder · Yotta Labs.
Published 8 November 2026. Every figure above was checked against the reporting linked here on the same day.