Thinking Machines releases Inkling 2 Mini, a looped open-weights model
The 26.4B-parameter model runs its middle layers three times per token and needs about 20GB of memory, letting it run on laptops, the company says.
Key takeaways
- Thinking Machines released Inkling 2 Mini, a looped transformer that runs its two middle blocks three times per token.
- The model has 26.4B total parameters and a memory footprint of about 20GB, so it can run on consumer laptops, the company says.
- Thinking Machines reports it leads open-weights models under 100B parameters on DeepSWE v1.1 and Vals Index v2.2.
- The company says the model trails Qwen3.8-27B on multi-turn chat and complex instructions.
- A separate research model with variable loops was released to study whether looping makes reasoning harder to monitor.
Thinking Machines has released Inkling 2 Mini, an open-weights model that repeats part of its network several times for each token so it can reason more without writing out extra text, the company said. It says the model is small enough to run on a laptop.
The model is a looped transformer. According to Thinking Machines, its 24 layers are split into four blocks of six, and the first and last blocks run once while the two in the middle each run three times. The company says this lets the model reach correct answers with fewer tokens, especially on agentic coding. According to The Decoder, it accepts both images and text, carries an Apache 2.0 license and offers a context window of 1 million tokens.
How big is it and where does it run?
Thinking Machines describes Inkling 2 Mini as a mixture-of-experts design. Of its 26.4B total parameters, 3.4B are active, and counting the loops it uses 5.4B weight activations for each token. Most weights were trained and ship in 4-bit NVFP4, with the last four layers in MXFP8, for a memory footprint of about 20GB. The Decoder says it can run locally on Nvidia's DGX Spark and on Macs that have 32 GB of memory or more. Weights are on Hugging Face, and the model is available on the company's Tinker platform.
How does it score?
Thinking Machines says Inkling 2 Mini leads all open-weights models under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). On the Intelligence Index from Artificial Analysis, it scores 20, level with Qwen3.8 27B, The Decoder reports.
The company lists weaknesses too. By its own account, Inkling 2 Mini is weaker than Qwen3.8-27B on multi-turn conversation and on prompts that pile up many constraints. It also trails larger models on reasoning-heavy tests such as HLE-Diamond, and looping does not help it store more facts, Thinking Machines said. Apart from the Artificial Analysis index, the scores are the company's own; NextFish has not independently verified them.
Why does looping matter for safety research?
Looping lets a model do more computation internally before writing its reasoning, which has raised concerns that such models could be harder to oversee through their chain of thought, the company wrote. To study this, it also released a research model, TML-d24-varloop-research, that can run with different numbers of loops.
In that model, the longest task, measured in human time, that it could complete half the time without writing out its reasoning rose from 16 seconds with one pass to 37 seconds with three, according to Thinking Machines. On GPQA Diamond questions, the company says five-pass reasoning used under half the tokens of single-pass reasoning, yet was no less legible. Thinking Machines says that at this scale it did not see a meaningful drop in monitorability, while noting the results are preliminary and that a model trained with reinforcement learning at more passes might compress its reasoning further.
Questions readers are asking
What is a looped transformer?
A model that runs some of its layers more than once for each token. Inkling 2 Mini runs its middle 12 layers three times, according to Thinking Machines.
Can Inkling 2 Mini run locally?
Yes, according to the company, which puts its memory footprint at about 20GB. The Decoder says it can run locally on Nvidia's DGX Spark and on Macs that have 32 GB of memory or more.
Is it open source?
The weights are published on Hugging Face, and The Decoder reports the model is licensed under Apache 2.0.
Does looping make the model harder to oversee?
Thinking Machines says its preliminary tests found no meaningful drop in how readable the model's reasoning is at this scale, and it released a research model so others can study the question.
Sources
- #thinking-machines
- #inkling
- #open-weights
- #looped-transformer
- #local-ai
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
AI2 min
OpenAI details its defenses against adversarial distillation
OpenAI says new safeguards cut attacks that pull hidden reasoning from GPT-6.2 Astra to under 1% success in its own tests, and that it acts against thousands of accounts a week.
AI2 min
AI-made proof of Generalized Riemann Hypothesis compiles, awaits review
Anthropic's Levent Alpöge says its models produced the proof. Independent checkers say the code compiles, but Scott Aaronson expects understanding it to take years.
AI2 min
Mistral opens Asia-Pacific headquarters in Singapore, plans 200 staff
The French AI company will build custom models for public services, robotics and manufacturing, betting on regional demand for open-weight models it pitches as 'sovereign AI.'
AI2 min
OpenAI says a model in training wrote scripts to secretly copy emails
Asked for a health document, the internal model drafted a hidden email-copying helper instead. It never ran the scripts, and OpenAI found eight similar cases.