Overview
Inkling 2 Mini is the smallest model in Thinking Machines Lab's second Inkling generation, released on November 20, 2026. It is a Mixture-of-Experts looped transformer with 26.4 billion total and 3.4 billion active parameters. The Hugging Face model card lists text and image input, text output, a context window of up to 1M tokens and the Apache 2.0 license.
The model stores 24 layers in four blocks of six. The two middle blocks run three times per token, so the model applies 48 layers from 24 stored ones; counting the loops, the lab gives 5.4B weight activations per token. Thinking Machines reports that looping the middle layers gave larger gains than looping the last layers on 27 of 28 evaluations.
Training began with 384 experts per layer. Two-thirds were pruned during midtraining, leaving 128, and the model was healed with distillation on 1T tokens. Most weights are NVFP4 and the last four layers are MXFP8, the precisions the model was trained in, and the lab gives a memory footprint of about 20GB.
| Released | 2026-11-20 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 26.4B total · 3.4B active |
| Context | 1M |
| Architecture | Mixture-of-Experts looped transformer: 24 stored layers in four blocks of six, with the two middle blocks run three times per token |
| Modalities | Text, Vision |
| Status | Generally available |
Strengths
- About 20GB of memory; the lab lists Ollama on Macs with at least 32GB of RAM
- Reported 79.2% on SWE-bench Verified, against 77.8% for Qwen3.8-27B
- Reported 82.7% on MCP Atlas and 79.0% on BrowseComp with compaction
- Up to 1M tokens of context with text and image input
Best for
- Reach for it to run a coding or agent model locally on a 32GB Mac or an NVIDIA DGX Spark.
- Reach for it to fine-tune a small open model on Tinker or with SkyRL, Miles or NVIDIA AutoModel.
- Reach for it to study looped transformers, alongside the TML-d24-varloop-research model released with it.
How to access
| Provider | Model ID |
|---|---|
| Tinker (Thinking Machines) ↗ | — |
| Hugging Face (open weights) ↗ | — |
Inkling — every version
The full lineage of the Inkling line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Inkling 2 Minicurrent | 2026-11-20 | 1M | Apache-2.0 |
| Inkling 2 Small | 2026-11-05 | 1M | Apache-2.0 |
| Inkling-Small | 2026-07-30 | 1M | Apache-2.0 |
| Inkling | 2026-07-15 | 1M | Apache-2.0 |
FAQ
What is Inkling 2 Mini?
Inkling 2 Mini is Thinking Machines Lab's smallest second-generation model, released on November 20, 2026. It is a 26.4B total / 3.4B active looped Mixture-of-Experts model with text and image input, up to 1M tokens of context and Apache-2.0 weights on Hugging Face.
What does looping mean in Inkling 2 Mini?
Inkling 2 Mini runs its two middle blocks of six layers three times for every token, so 24 stored layers act as 48. Thinking Machines Lab says the middle layers carry more abstract reasoning, and looping them beat looping the last layers on 27 of 28 evaluations.
How much memory does Inkling 2 Mini need?
Thinking Machines Lab gives Inkling 2 Mini a memory footprint of about 20GB, with most weights in NVFP4. The lab lists Ollama support on Macs with at least 32GB of RAM, GGUF builds through Unsloth and llama.cpp, and NVIDIA DGX Spark.

