Inkling 2 Mini: A 26B Looped AI Model That Runs on Your Laptop

Amir Farid
7 min read

Inkling 2 Mini is a 26.4-billion-parameter open-weight AI model from Thinking Machines Lab that runs on a laptop — and, per the lab's own benchmark reports, codes about as well as models several times its size. Released on 20 November 2026 under an Apache 2.0 licence, it runs a capable coding and agentic model locally, without sending your code or data to a cloud provider. For a small business, that changes the maths on private AI agents: no per-token bills, no data leaving the office, and an agent that keeps working when the internet does not.
What Inkling 2 Mini actually is
Inkling 2 Mini is the small, local sibling of Inkling 2 Small, the 594-billion-parameter model released earlier in November — see our Inkling 2 Small guide. Where the Small is a data-centre model, the Mini fits on one machine.
The looped transformer trick
Inkling 2 Mini is a looped transformer: it has 24 physical decoder layers, but the middle 12 run three times for every token, giving 48 logical layer visits from 24 layers of weights. It is a mixture-of-experts model too — 128 experts per layer, 8 active per token — with text and image input and 1M-token context. The lab's numbers: 26.4 billion total parameters, 3.4 billion active per token, 5.4 billion weight activations per token once looping is counted. A normal model gets deeper by adding layers — more weights to store and load. A looped model reuses the same layers, so it behaves like a much deeper model while fitting in about 20GB.
Shipped ready to run, not ready to quantise
The model ships natively in NVFP4, with the last four layers in MXFP8. Most small models only reach a laptop after compression, which usually costs accuracy; here there is no extra quantisation step and, according to the lab, no extra quantisation loss. It ships with GGUF builds (IQ4_NL and Q8_0) for llama.cpp, MLX for Apple silicon, SGLang from day zero, vLLM from launch day and Transformers support. Tool calling works, a three-layer speculative-decoding drafter called DSpark speeds up generation, and reasoning effort runs across seven levels, default high (0.9). Full details are on the Inkling 2 Mini model page on Hugging Face.
How good is it at coding? The benchmarks, with a health warning

Every number here comes from Thinking Machines Lab's own reports, published via Hugging Face — the lab's claims about its own model, not independently verified results.
The headline scores
The lab reports 20.1 on the AA Index v5 (AAv5), 48.2% on DeepSWE v1.1 and 77.9% on the hard split of SWE-bench Pro v2. It leads FrontierCode v2.0, according to the lab, and ties Qwen3.8-27B on Terminal-Bench 4.0. Hugging Face reports it leads in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL — see the Hugging Face announcement for the full results.
Cheaper and faster per task, the lab says
The most interesting claim is not a score but a cost. The lab reports Inkling 2 Mini is roughly six times cheaper and four times faster per task than Qwen3.8 27B — $0.59 against $3.53, and 9.0 minutes against 38.2 minutes per Intelligence Index task. If those figures hold up, a local model stops being a compromise and becomes the cheaper way to run coding agents. Our GDPx agent benchmark guide explains why task-completion benchmarks matter more than raw knowledge tests. The field is moving fast — Kimi K3.1 Pro, Qwen4 Max and Claude Fable 5.2 all target the same job — but Inkling 2 Mini's pitch is different: not the biggest model, a capable one you can own outright.
What hardware do you need to run it?
Official targets: Apple silicon with 32GB or more of unified memory — demoed offline on a MacBook Pro with the M5 Pro chip and 48GB — a single NVIDIA DGX Spark, or a GPU setup with 32GB or more of VRAM. At roughly 20GB, that means a well-specced laptop, not a server rack.
Memory first: 32GB is the floor the lab names, and 16GB will struggle or fail to load the model. Apple silicon is the smoothest path — MLX and GGUF builds make a modern MacBook the easiest local setup — while Windows and Linux work via llama.cpp, vLLM and SGLang.
One warning: the lab's research sibling, TML-d24-varloop-research, varies the loop count from one to five passes at inference. The lab and Hugging Face warn it can behave unpredictably away from three passes and it is explicitly not for production — stick to the standard model.
What local AI actually buys your business
Your data stays in the building
Code, customer records and internal documents never leave your office. If you handle client data under contract, that removes a whole category of compliance questions — a private AI agent running locally is an easier sell to a cautious client than one sending everything to a third-party API.
No per-token bills
Cloud models charge by the token, and agentic coding burns tokens fast. A local model costs the hardware once and electricity after that, so a small team running agents all day can break even quickly — our AI agent development cost guide for the UK has the typical cloud running costs.
It works offline, and you control the version
The lab's MacBook demo ran fully offline — handy on site visits, on trains, and in factories with patchy Wi-Fi. And a local model is a file you own: under its Apache 2.0 licence it behaves the same tomorrow as today, while cloud providers update, deprecate and reprice models without asking.
Where the cloud still wins
The biggest cloud models still out-reason a 26-billion-parameter local model on the hardest problems — if a task needs the deepest reasoning you can buy, the cloud wins. For routine coding the gap is much smaller, and that is where most business work sits. Scale is the second gap: one laptop serves one or two people at a time; cloud APIs serve a whole team instantly. Maintenance is the third — updates, drivers and troubleshooting land on you or your IT support. If nobody in your business wants to own a machine, the cloud's convenience is worth paying for. Our AI automation team can help you weigh the trade-off.
Putting Inkling 2 Mini to work: a practical starting point
Adopt it with a small pilot. A sensible order:
- Pick one machine — a MacBook Pro with 32GB or more, or a workstation with a 32GB GPU.
- Pick one job: code review, an internal documentation agent, or a RAG chatbot over your own documents.
- Measure it for two weeks — how often the agent completes tasks unaided, versus your cloud setup on cost and speed — then decide what stays local.
If you want a local agent wired into your real systems — CRM, inbox, WhatsApp — that is a build job, not a download. Our custom software development and SaaS development teams build private agents; our guide to local personal agents with n8n shows how local tooling fits together. If voice is part of the plan, see how to build an AI voice agent.
Frequently asked questions
Can Inkling 2 Mini really run on a laptop?
Yes, on the right laptop: Apple silicon with 32GB or more of unified memory, a single NVIDIA DGX Spark, or 32GB or more of VRAM. It was demoed offline on a MacBook Pro with the M5 Pro chip and 48GB. A 16GB machine will not be enough.
How does it compare with the biggest cloud models?
On the lab's own reported benchmarks it is competitive for its size — 48.2% on DeepSWE v1.1 and 77.9% on the hard split of SWE-bench Pro v2. It will not out-reason the largest frontier models on the hardest tasks, but for routine coding the lab's figures suggest it is close enough.
Should my business switch to local AI agents now?
Start with a pilot. If your work involves sensitive code or documents, heavy daily agent use or unreliable internet, it is worth testing now. If you need maximum reasoning power or serve many users at once, cloud models still make more sense. For an honest assessment, book a free AI consultation with Coheser; see also our AI agents for small business automation guide and chatbot vs AI agent explainer.