Last updated: November 21, 2026
Inkling 2 Mini is an Apache 2.0 open-weights model that Thinking Machines Lab released on November 20, 2026. It has 26.4B parameters, loops its middle layers three times per token, ships as a 20GB checkpoint, and runs on Macs with 32GB of memory through Ollama, on DGX Spark, or on Tinker.
Contents
- What is new in Inkling 2 Mini
- Specifications at a glance
- How looping works
- How to run it locally
- Serving with vLLM and other runtimes
- Tinker access and pricing
- Benchmarks and comparison
- Recommended workflows and settings
- Known limitations
- Frequently asked questions
What is new in Inkling 2 Mini
Inkling 2 Mini follows Inkling 2 Small on Tinker and the later Inkling 2 Small open weights release. It is a smaller model that Thinking Machines describes as small enough to run on a laptop.
According to the launch announcement, the headline change is architecture: it is a looped transformer, the first Thinking Machines has released. A research variant with a variable number of loops, TML-d24-varloop-research, was released alongside it for interpretability work.
The weights were trained and are shipped in the same low-bit precision, NVFP4 for 20 layers and MXFP8 for the last 4, so there is no conversion loss from quantization.
Specifications at a glance
Figures below come from the official model card, the Ollama library page and the Tinker price list.
| Item | Inkling 2 Mini |
|---|---|
| Release date and license | November 20, 2026; Apache 2.0 |
| Parameters | 26.4B total, 3.4B active, about 5.4B effective active per token |
| Architecture | 24-layer MoE decoder, 8 of 128 experts per token, hybrid local and global attention |
| Inputs and outputs | Text and images in (40px to 4096px per side recommended); text out |
| Context window | Up to 1M tokens (model card); 64K and 256K tiers on Tinker |
| Checkpoint | About 20GB, NVFP4 plus MXFP8 |
| Supported hardware | 1x NVIDIA DGX Spark; Apple silicon with 32GB+; other GPUs with 32GB+ VRAM recommended |
| Hugging Face repo | thinkingmachines/Inkling-2-Mini |
| Ollama tags | inkling-2-mini, :26b-a5b, :26b-a5b-nvfp4 (MLX) |
How looping works
The model has 24 layers in four blocks of six. The first and last blocks run once; the two middle blocks run three times with shared weights. That gives 48 layer applications per token from 24 stored layers, and each extra pass adds about 1.0B effective active parameters.
Thinking Machines says looping lets the model reason more without writing more tokens. In its tests, an unpruned non-looping model cost 7.5% less to train but needed 140% more inference compute to match the looped version on 512K-token agentic traces. The practical effect: compute per token is higher than the 3.4B active figure suggests, while knowledge recall tracks total parameter count and does not improve with more loops.
How to run it locally
For a Mac with 32GB or more of unified memory, Ollama is the fastest path:
- Install or update Ollama on your Mac.
- Check free memory and disk. Every listed tag is about 20GB.
- Run
ollama run inkling-2-minito pull the default tag and open a chat. Useinkling-2-mini:26b-a5b-nvfp4for the MLX build. - Call it from code through the local API at
http://localhost:11434/api/chatwith"model": "inkling-2-mini". - Connect a coding agent with
ollama launch claude --model inkling-2-mini; the library page also lists OpenCode, Hermes Agent and OpenClaw. - Set limits on output tokens and agent turns before long sessions, as the model card recommends.
Ollama tags the model with vision, tools and thinking capabilities. Quantized GGUF builds are available from Unsloth and llama.cpp; our Qwen3.8-27B GGUF guide explains how Unsloth quant levels trade memory for quality.
Serving with vLLM and other runtimes
On NVIDIA hardware, the Hugging Face model page shows an OpenAI-compatible server started with vllm serve "thinkingmachines/Inkling-2-Mini"; SGLang uses python3 -m sglang.launch_server. The model card also lists Unsloth, NVIDIA Dynamo, TokenSpeed and Llama App, with Exo Labs on Apple silicon. For recent vLLM changes, see our vLLM 0.33 guide.
The chat template accepts a reasoning_effort setting: none, minimal, low, medium, high, xhigh or max, or a number from 0.0 to 0.99. If you omit it, the template uses 0.9, the same value as high. The model card notes that the lowest setting does not fully turn off reasoning.
For RL fine-tuning, Thinking Machines lists Miles, SkyRL and NVIDIA NeMo AutoModel. DGX Spark owners may also find our Nemotron for Science guide useful.
Tinker access and pricing
Tinker provides API access for sampling and fine-tuning, and the Tinker Playground offers chat-only access. The price list, in US dollars per million tokens, shows:
| Model and context | Prefill | Cached prefill | Sample | Train |
|---|---|---|---|---|
| Inkling 2 Mini, 64K | $0.15 | $0.015 | $0.50 | $0.75 |
| Inkling 2 Mini, 256K | $0.15 | $0.015 | $0.50 | $1.50 |
| Inkling 2 Small, 64K | $0.30 | $0.03 | $1.00 | $1.50 |
Checkpoint storage is $0.10 per GB per month. Tinker’s serverless inference beta lists Inkling 2 Small, Inkling and Inkling-Small, but not Inkling 2 Mini, at the time of writing.
Benchmarks and comparison
Selected scores reported by Thinking Machines, compared with Qwen3.8-27B:
- DeepSWE v1.1: 48.2%, versus 42.2% for Qwen3.8-27B.
- SWE-bench Verified (bash only): 79.2%, versus 77.8%.
- MCP Atlas: 82.7%, versus 77.3%.
- BrowseComp with compaction: 79.0%, versus 61.6%.
- Terminal Bench v2.1: 71.6%, slightly behind Qwen3.8-27B at 73.0%.
- ScreenSpot Pro and OmniDocBench v1.6: 76.5% and 92.8%.
The announcement notes that evaluation setups vary by model.
Recommended workflows and settings
- Offline coding agent: pair the Ollama build with Claude Code or OpenCode, run it in a sandbox, and ask it to plan and wait for approval before editing or deleting files.
- Search-backed research: give it a search or MCP tool, since its tool use is stronger than its built-in factual recall.
- Screen and document reading: use image inputs for screenshots, charts and PDFs.
- Cost control: use high effort for debugging and low or minimal for simple extraction, and cap output tokens.
Known limitations
- Repetitive reasoning: it can get stuck re-checking its work, especially when debugging.
- Agent reliability: it may delete or overwrite files without confirmation and may misreport results.
- Knowledge: it is weaker than Inkling 2 Small on factual recall and more likely to be confidently wrong.
- Chat and constraints: it scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B.
- Languages: quality can drop outside English.
- Local speed: looped model support in local runtimes is new, so speed varies.
Frequently asked questions
What is Inkling 2 Mini?
Inkling 2 Mini is an open-weights model from Thinking Machines Lab, released on November 20, 2026 under Apache 2.0. It is a looped Mixture-of-Experts model with 26.4B total parameters and 3.4B active, accepts text and images, and is built for coding, agents and local use.
Can I run Inkling 2 Mini on a Mac?
Yes. Thinking Machines supports Apple silicon Macs with 32GB or more of memory through Ollama and Exo Labs. The checkpoint is about 20GB. On a supported Mac, run ollama run inkling-2-mini to download and start the model.
What is the Inkling 2 Mini context window?
The model card states a context window of up to 1M tokens, and Ollama lists 1M for its tags. On Tinker, Thinking Machines sells the model at 64K and 256K context lengths, with different training rates for each.
How much does Inkling 2 Mini cost?
The weights are free under Apache 2.0. On Tinker, the price list shows $0.15 per million prefill tokens, $0.015 cached prefill, $0.50 sampling, and $0.75 training at 64K context or $1.50 training at 256K context.
What does looping mean in Inkling 2 Mini?
Of its 24 layers, the middle 12 run three times for every token with shared weights. That gives 48 layer applications per token and about 5.4B effective active parameters, so the model does more reasoning per token without writing more tokens.
How is Inkling 2 Mini different from Inkling 2 Small?
Inkling 2 Mini is the smaller, local-first model. Thinking Machines says it is less capable than Inkling 2 Small across most evaluations, especially factual recall, but its Tinker rates are half of Inkling 2 Small’s at 64K context.
Sources
- Thinking Machines Lab: Introducing Inkling 2 Mini (November 20, 2026)
- Thinking Machines Lab: Inkling 2 Mini model card
- Hugging Face: thinkingmachines/Inkling-2-Mini
- Ollama library: inkling-2-mini
- Thinking Machines Lab: Tinker pricing
Image: Thinking Machines Lab, Inkling 2 Mini announcement cover (thinkingmachines.ai), resized; used for editorial purposes with credit. FreeAnswers is not affiliated with Thinking Machines Lab.