Inkling 2 Mini runs three times over its middle layers to fit a laptop
Thinking Machines released Inkling 2 Mini, an open-weight model that repeats its middle layers to get more compute from 3.4B active parameters, and it now appears on OpenRouter.

Thinking Machines Lab has released Inkling 2 Mini, an open-weight multimodal model with 26.4B total parameters, of which 3.4B are active per token. Its middle layers run three times with shared weights, and the model now appears on OpenRouter.
A looped design meant for laptops
According to the OpenRouter listing, Inkling 2 Mini is a mixture-of-experts looped transformer. Because the middle layers are reused three times, the model does more computation per token than its active parameter count implies, while staying small enough for a laptop or a single workstation GPU. Thinking Machines’ own release post, dated November 20, puts the active count at 5.4B when the looping is included, and gives a memory footprint of about 20GB. The weights are quantized to NVFP4, with the last four layers in MXFP8.

It is the smallest model of the second Inkling generation, below Inkling 2 Small. The listing says it accepts text and images and returns text, with a context window of 1,048,576 tokens. The company recommends it for agents that call tools, code assistants, conversational bots and search-backed question answering, and says it covers English along with a broad range of other languages.
Where you can actually run it
The weights are on Hugging Face under thinkingmachines/Inkling-2-Mini, and the company also published a variable-loop research variant. According to the release post, the model is available on Tinker, the company’s platform for inference and fine-tuning, and runs through Ollama on Macs with 32GB or more of RAM. Other supported runtimes include SGLang, vLLM, Unsloth and NVIDIA Dynamo, plus GGUF builds for llama.cpp.

The OpenRouter route is less settled. The page shows no standard public endpoint and no price, and says the model is not available through EU in-region routing. The API model ID is thinkingmachines/inkling-2-mini. The release post does not name a license, a price or a maximum output length, and OpenRouter is not mentioned in it, so readers should check the license on Hugging Face before building on it.
Small and open, but not alone in its class
In Thinking Machines’ own comparison table, Inkling 2 Mini is set against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. The company reports 79.2% on SWE-bench Verified (bash only), 82.7% on MCP Atlas, 79.0% on BrowseComp with compaction and 48.2% on DeepSWE v1.1, ahead of Qwen3.8-27B’s 42.2% on that last test. The picture is not one-sided: Qwen3.8-27B scores 73.0% on Terminal Bench v2.1 against 71.6% for Inkling 2 Mini, and the two are essentially level on HLE-Diamond, at 15.3% and 15.2%. The company also lists weaker results itself, such as 50.0% on MultiChallenge and 33.4% on ComplexConstraints. Most of these figures come from the company’s post and have not been independently verified, apart from rows attributed to Artificial Analysis and Vals.

The size choice stands out against recent open releases we have covered. Tencent’s Hy4 flash is a 220B model with 12B active parameters, and it costs $0.04 per million input tokens on OpenRouter while keeping a 1M context. Inkling 2 Mini offers the same 1M window with roughly a quarter of the active parameters and about an eighth of the total size. Tencent’s larger 770B Hy4 base sits at the other end of the open-weight range. Thinking Machines is betting that repeated layers can substitute for raw size, and that a model you can host yourself matters more than a hosted price.
That bet has a cost the source does not quantify: running the middle layers three times adds compute per token, so speed on a given laptop or GPU will depend on the runtime. The release post cites a Stanford-derived intelligence-per-watt score of 0.703 on an NVIDIA DGX Spark, against 0.430 for Qwen3.6-35B-A3B, but real-world latency and quality on local hardware remain to be shown by outside testers.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Tencent’s Hy4 appears on OpenRouter at half the preview’s price

Tencent opens the base weights of Hy4 flash, a 220B model with 12B active

GLM-5.3-Flash on MI300X: Switching parallelism doubles long-context throughput
