Inkling 2 Mini: Run Thinking Machines' Open Model Locally
Inkling 2 Mini is a 26.4B-parameter open-weight multimodal model from Thinking Machines Lab, runnable locally on a 32GB Mac or DGX Spark.

What is Inkling 2 Mini?
Inkling 2 Mini is an open-weight multimodal model from Thinking Machines Lab, released under Apache 2.0 on November 20, 2026. It takes text and image input and produces text output, and it’s built specifically to be small enough to run on local hardware rather than requiring a data center GPU cluster. The full checkpoint is about 20GB, distributed in mixed NVFP4 and MXFP8 precision, which is why it fits on a 32GB Mac or a single DGX Spark unit.
TL;DR
- Inkling 2 Mini has 26.4 billion total parameters but only 3.4 billion active per token, which is what makes local inference realistic on consumer-grade hardware.
- The architecture is a looped transformer: 24 layers of weights, but the middle 12 run three times, producing 48 effective layer applications and about 5.4B effective active parameters per token once the loop passes are counted.
- It supports a 1M token context window and takes both text and images as input, with images processed through a vision transformer into the same hidden space as text.
- Minimum recommended hardware is 32GB of unified memory or VRAM, officially validated on Apple silicon Macs and a single NVIDIA DGX Spark.
- The model is supported across a wide software stack including Ollama, vLLM, llama.cpp, SGLang, Unsloth (GGUF), Hugging Face transformers, and NVIDIA Dynamo.
- On coding-heavy benchmarks like SWEBench Verified and SWEBench Pro, Inkling 2 Mini posts scores competitive with or ahead of larger dense models such as Qwen3.8-27B and Nemotron 3.5 Super VL.
- It’s also accessible through Thinking Machines’ hosted Tinker platform, which offers both a playground for inference and an API for fine-tuning.
Build an onboarding app for 4,000 new hires.
Build ↵How does the looped transformer architecture work?
Inkling 2 Mini departs from a standard decoder-only stack. It has 24 layers of weights, but the middle 12 layers are reused: the model runs them three times in a loop with shared weights. That gives the effective depth of a 48-layer model while only storing the parameters for 24. Each layer also uses sparse Mixture-of-Experts feed-forward blocks, routing each token to 8 of 128 available experts, and attention mixes local and global layers to balance cost against long-context performance.
This looping is the reason the “active parameters” number isn’t static. The base active parameter count is 3.4B, but Thinking Machines notes that each additional pass through the looped layers adds roughly 1.0B in effective active parameters. With three passes, that works out to about 5.4B effective active parameters per token. In practice, this means the model does more computation per token than a typical MoE of similar on-disk size would suggest, trading some inference speed for reasoning depth without inflating the file size you have to download and store.
The model is natively multimodal rather than bolted together. Images go through a vision transformer encoder, and the resulting representations are projected into the same hidden space as text tokens before being handed to the decoder. Accepted image sizes range from 40px to 4096px per dimension.
What hardware do you need to run it locally?
The official requirements are modest by frontier-model standards:
- A single NVIDIA DGX Spark
- Any Apple silicon Mac with 32GB or more of unified memory (MacBook, Mac mini, or Mac Studio)
- Other GPUs or platforms with 32GB or more of VRAM, which Thinking Machines says should also work even though they aren’t the primary validated targets
The 20GB checkpoint size, combined with the NVFP4/MXFP8 mixed-precision format, is what keeps the memory footprint low enough for a single consumer-class device. This is a meaningfully different deployment story than most 26B-parameter dense models, which would typically need 50GB+ of VRAM just to hold weights at higher precision. Because only 3.4B parameters activate per token (before loop overhead), inference throughput on local hardware is also closer to what you’d expect from a small dense model than from a 26B one.
Which frameworks and tools support Inkling 2 Mini?
Thinking Machines shipped Inkling 2 Mini with support across most of the common local-inference and serving stacks developers already use:
- Hugging Face transformers, for direct Python integration
- vLLM and SGLang, for higher-throughput serving
- llama.cpp and Unsloth, both supporting quantized GGUF versions for lighter-weight local runs
- Ollama and Exo Labs, specifically called out for Apple silicon
- NVIDIA Dynamo and TokenSpeed, for production-style serving setups
- Llama App, as a packaged client
For fine-tuning or post-training, the model also integrates with Miles and SkyRL, two reinforcement-learning and post-training frameworks. Developers who want hosted access without managing local infrastructure can use the Tinker playground for inference or the Tinker Cookbook for programmatic fine-tuning access, which requires installing the tml-renderers package alongside it.
This breadth of support matters more than it might seem. A lot of open-weight releases land with only one or two inference paths working cleanly at launch, leaving GGUF conversions and quantization support to community effort over the following weeks. Inkling 2 Mini shipping with Ollama, vLLM, and llama.cpp support (including GGUF) from day one lowers the barrier for anyone who wants to try it on a laptop the same day it’s released.
How does it perform compared to similarly sized models?
Thinking Machines’ own evaluation table compares Inkling 2 Mini against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL. The comparison is notable because Inkling 2 Mini runs with only 5 billion effective active parameters against dense competitors activating 27 to 31 billion.
On coding and agentic benchmarks, Inkling 2 Mini is frequently at or near the top of its comparison group:
- SWEBench Verified (bash-only agent): 79.2%, the highest score in the group
- SWEBench Pro v2 (Hard): 77.9%, ahead of Qwen3.8-27B’s 73.8%
- SWEBench Multilingual: 73.7%, the top score listed
- DeepSWE v1.1: 48.2%, versus 42.2% for the next-best model
- Terminal Bench v2.1: 71.6%, slightly behind Qwen3.8-27B’s 73.0%
On the AA Index (v5.0), Inkling 2 Mini ties Qwen3.8-27B at 20.1, both well ahead of Gemma 4 31B (4.4) and Muse Glimmer (6.3). On the Vals Index (v2.2), it leads the group at 46.5%.
The pattern across the table is consistent: a model with a fraction of the active compute of its competitors matching or beating them on agentic coding tasks. That’s the practical payoff of the Mixture-of-Experts plus looped-layer design, more effective reasoning per token without the full memory and compute cost of a dense model that size.
Is Inkling 2 Mini worth running locally?
For developers who want a capable coding and agentic model without renting cloud GPU time, Inkling 2 Mini is a reasonable candidate. The combination of strong SWEBench-style scores, a 1M token context window, Apache 2.0 licensing, and broad framework support (particularly Ollama and llama.cpp on Apple silicon) makes it accessible to anyone with a modern Mac with 32GB of memory, which is a common configuration for developer laptops already.
The tradeoffs are the ones that come with any looped or MoE architecture: effective compute per token is higher than the raw “3.4B active” headline number suggests, so actual inference speed on constrained hardware should be benchmarked against your specific workload rather than assumed from parameter count alone. It’s also a very new release (November 2026), so community tooling like quantization quality for GGUF variants and third-party fine-tunes will likely mature over the following weeks.
Frequently Asked Questions
How many parameters does Inkling 2 Mini have?
It has 26.4 billion total parameters, with 3.4 billion active per token at baseline. Because the model loops its middle 12 layers three times, the effective active parameter count during inference is approximately 5.4 billion.
Can Inkling 2 Mini run on a MacBook?
Yes, on any Apple silicon Mac with 32GB or more of unified memory, including MacBooks, Mac minis, and Mac Studios. It’s supported through Ollama, Exo Labs, llama.cpp, and Unsloth’s GGUF builds.
What frameworks support Inkling 2 Mini?
Hugging Face transformers, vLLM, SGLang, llama.cpp, Unsloth, Ollama, Exo Labs, NVIDIA Dynamo, TokenSpeed, and the Llama App. Fine-tuning is supported through Miles, SkyRL, and Thinking Machines’ own Tinker Cookbook.
What’s the context window on Inkling 2 Mini?

What you'll learn · Dec 2
- 01How enterprises really build with AI
- 02What full-stack means and why it's critical for business software
- 03Prompting that holds up in production
- 04Build and deploy your own app, live
Up to 1 million tokens, which is large enough for long documents, extended codebases, or multi-turn agentic sessions without aggressive truncation.
Is Inkling 2 Mini free to use commercially?
It’s released under the Apache 2.0 license, which permits commercial use, modification, and redistribution, subject to Thinking Machines’ published Model Acceptable Use Policy.



