Thinking Machines releases open 20GB agentic model that runs on a laptop
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, with open weights under apache-2.0.

· Last verified: 21 Nov 2026 (Thinking Machines Lab, Hugging Face)

Summary
- Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, with open weights under apache-2.0.
- The 26.4B-parameter model ships as a single 20GB checkpoint running on one DGX Spark or 32GB Apple silicon.
- Local execution lets firms run coding and research agents on their own hardware instead of the cloud.
The latest
A 20GB agentic model that fits on a laptop is now openly downloadable. Thinking Machines Lab published Inkling 2 Mini on November 20, 2026, releasing the weights on Hugging Face under an apache-2.0 license alongside an acceptable use policy. The company described it as a fast, powerful model small enough to run on a personal machine, and said it leads every open-weight model under 100B parameters on two benchmarks.
Details
- The architecture: According to the model card, Inkling 2 Mini is a looped multimodal autoregressive transformer with 26.4B total parameters and 3.4B active, rising to 5.4B effectively active per token. Its 24 layers include 12 middle layers executed three times with shared weights, producing 48 layer applications per token.
- Routing and inputs: Each token is routed to 8 of 128 experts, the card states. Inputs are text and images, with images encoded through a vision transformer, while outputs are text only. The context window extends to 1M tokens, and the checkpoint uses mixed NVFP4 and MXFP8 precision.
- The hardware bar: The company said the model runs on a single NVIDIA DGX Spark, or Apple silicon machines with 32GB of memory or more, covering MacBook, Mac mini and Studio. For other hardware it recommends 32GB of VRAM or above, and demonstrated offline use on a 48GB M5 Pro MacBook Pro.
- How it was built: The blog post said the model derives from the company's predictive scaling ladder, with experts cut from 384 to 128 per layer before loops were introduced. Training after the reduction used 1T tokens distilled from a larger internal Inkling model, then supervised fine-tuning and reinforcement learning totaling 2.2T tokens over 2,000 steps.
- The coding scores: Against Qwen3.8-27B, Inkling 2 Mini scored 79.2% versus 77.8% on SWEBench Verified, 77.9% versus 73.8% on SWEBench Pro v2 Hard, and 48.2% versus 42.2% on DeepSWE v1.1. It trailed on Terminal Bench v2.1 at 71.6% against 73.0%.
- The agentic gaps: The widest margins came on agentic and browsing tasks: 79.0% versus 61.6% on BrowseComp, 49.6% versus 33.4% on JobBench, and 82.7% versus 77.3% on MCP Atlas Public. The model scored 46.5% on Vals Index v2.2 and 92.8% against 91.5% on OmniDocBench v1.6.
- Where it loses: Qwen3.8-27B outperformed it on MMMU Pro, 76.3% to 74.9%, and on AutomationBench-AA, 48.2% to 41.8%. The company said its model leads open-weight models under 100B parameters on DeepSWE v1.1 and Vals Index v2.2, trailing Qwen3.8-27B alone on some agentic coding benchmarks.
- Safety numbers: The card reports 89.5% on the adversarial FORTRESS evaluation and 97.8% on StrongREJECT. On two internal measures where lower is better, it logged 9.8% for misbehavior and 10.2% for destructive actions.
- Stated limits: The card says Inkling 2 Mini is less capable than Inkling 2 Small across most evaluations, with the largest gaps in information retrieval. A research model, TML-d24-varloop-research, shipped alongside it with a variable loop count of 1 to 5 passes.
- Where to run it: The model is available on Tinker for inference and fine-tuning, with the playground limited to inference. Listed inference frameworks include vLLM, SGLang, Hugging Face, TokenSpeed, NVIDIA Dynamo, Unsloth with quantized GGUF, llama.cpp, Llama App, Ollama and Exo Labs. Post-training frameworks are Miles and SkyRL. No price was disclosed.
Between the lines
The benchmark spread points to a specialized bet rather than a general one. The largest margins over Qwen3.8-27B cluster in browsing, job and tool-use tasks, while the losses fall on multimodal understanding and automation. Paired with a 20GB checkpoint that runs offline on consumer Apple hardware, the design reads as optimized for local agents doing work, not for broad reasoning.
What's next
Watch whether independent evaluations reproduce the DeepSWE v1.1 and BrowseComp margins, whether Thinking Machines Lab publishes pricing for Tinker access, and how quickly quantized builds appear across llama.cpp and Ollama.