🤖 Thinking Machines released Inkling 2 Mini on November 20, 2026: an open-weights model with 26 billion parameters and a file of roughly 20GB that runs on a laptop with 32GB of memory. The weights ship under Apache 2.0 on Hugging Face, which makes the release a direct answer to the users who said Inkling 2 Small didn't fit on their hardware — answered in their own size class. The real claim, though, is not the small file: instead of growing its weights, the model runs the middle of its layers three times per token, so it thinks more while writing fewer tokens.
⏱️ 5 min read
📌 In This Article
- 📌 What Inkling 2 Mini is, and where it sits in the family
- 📌 The looped transformer: why the layers run three times
- 📌 What the benchmarks say, and which numbers are the lab's own
- 📌 Where you can run the model
- 📌 The research variant: what deepening the loop buys you
- 📌 Murat Karakaya Academy verdict
- 📌 References
1. What Inkling 2 Mini is, and where it sits in the family
🤖 Mira Murati's Thinking Machines Lab announced the model on November 20, 2026 [1]. Inkling 2 Mini is the smallest member of the family: 26.4B parameters in total, 3.4B of them active per token [1]. The company framed the launch from its own official account with the words "users found Inkling 2 Small (594B parameters) too large to run on their own hardware" — Mini is the direct product of that feedback [5][8].
📐 The family's size ladder today reads:
- 🔹 Inkling (July 15, 2026): 975B total, 41B active [6]
- 🔹 Inkling-Small: 276B total, 12B active [6]
- 🔹 Inkling 2 Small (November 5, 2026): 594B total, 17B active [5]
- 🔹 Inkling 2 Mini (November 20, 2026): 26.4B total, 3.4B active [1][6]
🖼️ The model takes text and image input and produces text; the context window covers up to 1 million tokens [2].
2. The looped transformer: why the layers run three times
🔁 The architecture rests on a design the lab calls a "looped transformer" [1]. Twenty-four physical layers are arranged in four blocks of six: the first and last block run once per token, while the two middle blocks run three times over the same weights. The result is a 24-layer model that behaves as if it had 48 logical layers [1][3].
⚙️ The rest of the architecture [1][2][3]:
- 🔹 MoE structure: each token is routed to 8 of 128 experts per layer.
- 🔹 The model was pruned from 384 experts down to 128 during training, then recovered its capacity on roughly 1 trillion tokens distilled from a larger internal Inkling model [8].
- 🔹 After pruning came supervised fine-tuning, then a reinforcement-learning (RL) run of 2.2 trillion tokens and 2,000 steps [1].
- 🔹 Most of the weights were trained — and released — at 4-bit NVFP4 precision, with the last four layers at 8-bit MXFP8; total size roughly 20GB [1].
🧊 So the model was not squeezed down after the fact; it was trained at 4-bit precision from the start. That is why running it locally costs you no quantization penalty on top [7].
3. What the benchmarks say, and which numbers are the lab's own
📊 Start with the numbers the lab itself published: 48.2% on DeepSWE v1.1 and 46.5% on the Vals Index v2.2, which they report as beating every open-weight model under 100B [1]. In the comparison table in the model card — run against Gemma 4 31B, Qwen3.8-27B, Nemotron 3.5 Super VL and Muse Glimmer — they lead in 32 of 48 tests; on the Artificial Analysis Intelligence Index v5 row of that table the score is 20.1, the same as Qwen3.8-27B [2][8].
⚠️ The lab also states an important limit on its own: the loop does not let the weights store more information; knowledge capacity depends mostly on total parameter count. Sure enough, on AA-Omniscience, the factuality measure, its score sits in line with models of its own size class [1]. What the loop does return is efficiency in search and tool use: on BrowseComp, DeepSearchQA and MCP Atlas it leads the open models in its size class [1].
⚡ On the energy side, in Stanford researchers' intelligence-per-joule measurement, Inkling 2 Mini scored 0.703 on an NVIDIA DGX Spark; in the same measurement Qwen3.6-35B-A3B got 0.430 and Gemma-4-26B-A4B 0.159 [1].
4. Where you can run the model
- 🔹 Open weights on Hugging Face, Apache 2.0 license [2].
- 🔹 Locally with a single command in Ollama:
ollama run inkling-2-mini— a 20GB package, 1M context, text and image input; 32GB or more of memory is recommended on the Apple silicon side [7][2]. - 🔹 vLLM, SGLang, Unsloth, llama.cpp (GGUF) and MLX support arrived on release day [1][3].
- 🔹 On the Tinker API: $0.15 per million input tokens ($0.015 from cache), $0.50 output and $0.75 training — exactly half of Inkling 2 Small [4].
✈️ The lab shared two demos as well: researchers ran the model on a MacBook Pro (M5 Pro, 48GB) in airplane mode mid-flight, and it stood up a playable flight simulator in a few minutes [1]. In the other, all 20 characters of Thinkyville, a pixel town, are driven by this same model on a single offline MacBook [1].
5. The research variant: what deepening the loop buys you
🔬 Alongside Mini, Thinking Machines released a research model whose loop count can be varied: TML-d24-varloop-research [1]. Where Mini loops a fixed three times, this variant runs anywhere between one and five passes. In the company's preliminary results, an experiment without chain-of-thought stretched its time horizon from 16 seconds on one pass to 37 seconds on three; on a sequential reasoning test, performance kept improving up to ten loops — twice the number it was trained on [1]. As the loops increase, the model's chain of thought gets shorter but stays readable, and the agent-sabotage tests showed no increase [1]. The lab stresses that it is publishing these models openly so researchers can examine looped architectures for themselves [1].
🎓 Murat Karakaya Academy verdict
🧭 Let's keep the data and our opinion apart. What the data shows: this launch moves the argument from "bigger model" to "more compute per token." The looped architecture is no marketing ornament — the lab backs its claim with a monitorability study and an open research variant. Being able to tune how much a model thinks, independent of how many tokens it writes, and keeping the chain of thought readable while doing it, is a design choice the agent era should take seriously.
⚖️ A hard note on the quality of the numbers: most of the benchmarks on offer are the lab's own measurements. On the independent side, the clearest reading we have is a tie with Qwen3.8-27B on the Artificial Analysis index — not a clean win. Then there is the lab's own admission: the loop adds no knowledge; the model's encyclopedic capacity does not reach past its size class.
🎯 Our opinion: Inkling 2 Mini is the strongest candidate in its class for local agent work today, but the "best in class" claim will only mean something once independent tests multiply. The practical side is already clear: a 32GB laptop can carry a serious agentic coding load, and dropping Tinker to half of Small's price sets a reasonable balance between running locally and calling an API. For developers in Turkey the payoff sits right here: a strong coding agent is no longer a datacenter subscription — it is a laptop.
📚 References
- 🔗 thinkingmachines.ai — "Introducing Inkling 2 Mini" · Nov 20, 2026 · accessed Nov 22, 2026
- 🔗 huggingface.co — "Inkling 2 Mini model card" · accessed Nov 22, 2026
- 🔗 huggingface.co — "Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup" · Nov 20, 2026 · accessed Nov 22, 2026
- 🔗 tinker-docs.thinkingmachines.ai — "Models & Pricing" · accessed Nov 22, 2026
- 🔗 thinkingmachines.ai — "Introducing Inkling 2 Small" · Nov 5, 2026 · accessed Nov 22, 2026
- 🔗 thinkingmachines.ai — "Inkling — the model family" · accessed Nov 22, 2026
- 🔗 ollama.com — "inkling-2-mini library page" · accessed Nov 22, 2026
- 🔗 thinkingmachines.ai — "Inkling 2 Mini Model Card (official)" · accessed Nov 22, 2026
Hashtags: #MuratKarakayaAkademi #Inkling2Mini #ThinkingMachines #LoopedTransformer #AgenticCoding #OpenWeights