cd /news/artificial-intelligence/thinking-machines-releases-inkling-2… · home › topics › artificial-intelligence › article
[ARTICLE · art-173502] src=testingcatalog.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

Thinking Machines releases Inkling 2 Mini open-weight model

Thinking Machines Lab, led by Mira Murati, released Inkling 2 Mini, an open-weights Mixture-of-Experts model with 26.4B total parameters and 3.4B active, published on Hugging Face and served on Tinker for inference and fine-tuning. The looped transformer runs its middle 12 of 24 layers three times per token, ships most weights natively in NVFP4 with the last four layers in MXFP8 for a roughly 20GB footprint, and Thinking Machines reports it leads every open-weights model under 100B total parameters on DeepSWE v1.1 at 48.2% and Vals Index v2.2 at 46.5% while tying Qwen3.8-27B at 20.1 on the Artificial Analysis Index. The lab says the model trails Qwen3.8-27B on multi-turn chat and many-constraint instructions, and it follows the 594B-parameter Inkling 2 Small that users found too large to run themselves.

key takeaway

Inkling 2 Mini runs its middle 12 of 24 layers three times per token, and Thinking Machines reports the no-CoT time horizon grew from 16 seconds at one pass to 37 seconds at three while agentic sabotage scores did not trend higher.

by read3 min views5 publishedNov 20, 2026
Thinking Machines releases Inkling 2 Mini open-weight model
Image: Testingcatalog (auto-discovered)

Thinking Machines Lab has released Inkling 2 Mini, an open-weights model built to be fast, capable at agentic work, and small enough to run on a laptop. The weights for Inkling 2 Mini and a research-specific variable-loop variant are on Hugging Face, and the model is live on Tinker for inference and fine-tuning, with chat in the Tinker Playground.

Open-source inference is supported from day one on SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo and TokenSpeed, with RL fine-tuning on SkyRL, Miles and NVIDIA AutoModel.

Inkling 2 Mini is a looped transformer. Of its 24 layers, the middle 12 run three times per token, while the first and last six run once, letting the model apply more reasoning and reach the right answer with fewer tokens. It is a Mixture-of-Experts model with 26.4B total parameters, 3.4B active and 5.4B weight activations per token once looping is counted. Most weights were trained and are deployed natively in NVFP4, with the last four layers in MXFP8, so there is no quantization loss and the full footprint sits around 20GB. On a flight, the team ran it offline on a MacBook Pro and built a playable flight simulator in a few minutes.

According to Thinking Machines, Inkling 2 Mini leads every open-weights model under 100B total parameters on DeepSWE v1.1 with 48.2% and Vals Index v2.2 with 46.5%, and ties Qwen3.8-27B at 20.1 on the Artificial Analysis Index. It also tops its size class on BrowseComp, DeepSearchQA, and MCP Atlas, using search and tools to make up for the limited knowledge a small model can store. The company is upfront about the gaps: it trails Qwen3.8-27B on multi-turn chat and on instructions with many constraints.

Looping lets a model reason longer without writing its reasoning down, which raises questions about monitoring a model through its chain of thought. Thinking Machines studied this with the variable-loop model. Its no-CoT time horizon grew from 16 seconds at one pass to 37 seconds at three, the chain of thought got shorter but stayed legible, and scores on agentic sabotage tests did not trend higher as loops were added.

Thinking Machines Lab, led by Mira Murati, follows Inkling 2 Small, a 594B parameter model that some users found too big to run themselves. Mini is the answer to that feedback and one of the first looped models open to researchers. The lab is working with the National Deep Inference Fabric to bring it to interpretability researchers, and gave Goodfire early access to study how looped models work inside.

I think this release lands in a spot many local model fans were watching. We expected Qwen to drop a new SOTA model for local work, but we got Inkling instead, and it is going straight into my local testing.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @thinking machines lab 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/thinking-machines-re…] indexed:0 read:3min 2026-11-20 · —