Inkling 2 Mini: an open-weights coding model for your laptop
Thinking Machines released Inkling 2 Mini: 26.4B open weights, ~20GB, Apache 2.0, 1M context. What a laptop-class agent model means for your AI bill.
A coding agent model that fits on a laptop just shipped with open weights. Thinking Machines Lab released Inkling 2 Mini on November 20. It is a 26.4B-parameter model with a memory footprint of about 20GB, under an Apache 2.0 license. The lab says it leads every open-weights model under 100B parameters on agentic coding benchmarks. If part of your AI bill is agents doing routine work, some of that work may now run on hardware you already own.
What actually happened
From the lab's announcement and the Hugging Face model card:
- Size. Mixture-of-experts, 26.4B total parameters, 3.4B active per token. Looping brings it to 5.4B weight activations per token.
- The trick. It is a "looped" transformer: the middle 12 of its 24 layers run three times per token, so it reasons more without writing more tokens.
- Footprint. About 20GB, trained mostly in 4-bit NVFP4. The lab demoed it offline on a MacBook Pro with 48GB of memory.
- Specs. Text and image in, text out. Context window up to 1M tokens.
- License and access. Apache 2.0 weights on Hugging Face, with vLLM serving instructions on the card. Inference and fine-tuning are also on Thinking Machines' Tinker platform.
- Claimed scores. 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2. These are the lab's numbers.
- Admitted gaps. It trails Qwen3.8-27B on multi-turn chat (50.0% on MultiChallenge) and on instructions with many constraints. It stores fewer facts than bigger models, so it leans on search and tools.
This follows the larger Inkling 2 Small release.
Why it matters for your business
Most small teams send every AI task to a frontier API and pay per token. That is fine for hard work. It is overkill for the long tail: tagging tickets, cleaning product data, triaging code changes, drafting replies. A strong model that runs on one machine changes the math for those jobs, and keeps sensitive data off a third party's servers.
The catch: lab benchmarks are not your workload. Its weak spots, multi-turn chat and complex instructions, are exactly what a customer-facing bot needs. Use it for back-office agents first.
What we would do:
- Pick three repetitive tasks you pay an API for today.
- Run them on Inkling 2 Mini locally and on your current model. Score the outputs blind.
- Compare cost per task, including the hardware you already own.
- Route by result. Local for what passes, frontier API for what does not.
- Keep it swappable. Put an OpenAI-compatible endpoint in front so a better model is a config change.
Key takeaways
- Inkling 2 Mini: 26.4B open weights, about 20GB, Apache 2.0, 1M context
- The lab says it leads open models under 100B on agentic coding
- Benchmarks are lab-reported; it is weaker at multi-turn chat
- Back-office agent work is the best first test
- Score it on your own tasks before you move any spend
Paying frontier prices for routine agent work? We set up model routing that sends easy jobs to local or open models and hard jobs to the big APIs, with evals on your data. Run the numbers or see what we build.
Sources: Thinking Machines Lab, Hugging Face model card.
- #inkling-2-mini
- #open-weights
- #local-llm
- #thinking-machines
- #ai-coding
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Genie 4 in the Gemini API: interactive worlds at ~$6 a minute
Google's Genie 4 world model is in the Gemini API: 720p interactive worlds, 120-second sessions, about $0.10 a second, paid tier only. Billing stops on complete().
Read itMicrosoft To Do retires June 30, 2027: tasks don't move
Microsoft To Do apps retire June 30, 2027. Work tasks don't migrate to Planner, personal accounts get no replacement, and export isn't out yet. Plan the move now.
Read it