Thinking Machines released Inkling 2 Mini, a 26.4B-parameter open-weights model with only 3.4B active parameters that loops its middle layers three times per token for improved reasoning.
Thinking Machines has released Inkling 2 Mini, a 26.4-billion-parameter open-weights model that takes a structurally unusual approach to reasoning. Rather than making a single pass through its layers, the model reruns its middle layers three times per token — a design intended to improve multi-step reasoning without inflating the number of active parameters or compute costs at inference time.
The model is available for download and deployment under an open-weights license, putting it in the small class of publicly available models capable of multi-step reasoning without requiring proprietary infrastructure.
The Architecture
Inkling 2 Mini is a mixture-of-experts (MoE) model — a design where only a fraction of the model's total parameters activate on any given token. This keeps inference costs low relative to the total parameter count.
The key numbers:
- 26.4 billion total parameters
- 3.4 billion active parameters per token
- Middle layers execute three times per token (looped architecture)
- Open weights — no API, no access controls, fully downloadable
The MoE structure is not new — Mistral's Mixtral models and others use the same principle. What is unusual here is the looped middle-layer execution. Most transformer-based models process each layer exactly once per forward pass. Running the middle layers multiple times is closer in spirit to recurrent networks, which were dominant before transformers took over in 2017.
The company's claim is that this loop improves reasoning — specifically the kind of multi-step logic, planning, and inference tasks where current models often stumble.
Get this in your inbox.
Daily AI intelligence. Free. No spam.
Why Open Weights Matter Here
Proprietary reasoning models from OpenAI, Anthropic, and Google deliver strong performance, but they come with constraints: per-token API costs, rate limits, no fine-tuning access, and dependency on a third-party's infrastructure and pricing decisions.
An open-weights model that performs comparably on reasoning benchmarks changes the calculus for:
- Independent developers who cannot afford commercial API costs at scale
- Enterprises with data privacy requirements that prevent sending queries to external servers
- Researchers who need to study model internals, probe failure modes, or fine-tune for specialized domains
- On-device deployment scenarios where network latency or connectivity is a constraint
The 3.4B active-parameter footprint is small enough to run on a single high-end consumer GPU, which opens deployment to a much wider hardware base than larger models.
What Independent Benchmarks Will Tell Us
The looped architecture is an interesting architectural claim, but the proof is in benchmarks. The key questions independent evaluators will want to answer:
- Does the looped middle-layer execution actually improve scores on established reasoning benchmarks (ARC-AGI, GPQA, MATH) compared to comparable active-parameter models without the loop?
- Does the loop add measurable latency per token, and if so, how does the accuracy-latency tradeoff compare to alternatives?
- How does Inkling 2 Mini perform against Mistral, Llama, Qwen, and Phi at similar active-parameter budgets?
Until independent benchmarks arrive, Thinking Machines' architecture claims should be treated as promising hypotheses, not settled results.
What to Watch
Watch for community benchmark evaluations on Hugging Face and in AI research publications over the next few weeks. The looped-layer design is novel enough that if it shows consistent improvements, it could influence future open-model architecture choices across the field.
The release also adds competitive pressure to the open-weights ecosystem heading into a period where the gap between open and closed models has been slowly narrowing. Whether Inkling 2 Mini's architecture contributes to closing that gap further is the question the benchmarks will answer.
Source: AI Universe Journal, November 21, 2026
Did this help you understand AI better?
Your feedback helps us write more useful content.
Get tomorrow's AI briefing
Join readers who start their day with NexChron. Free, daily, no spam.