the short version
- Inkling 2 Mini has 26.4B total and 3.4B active parameters. It loops its middle 12 layers three times, which Thinking Machines counts as 5.4B weight activations per token.
- Most weights are in NVFP4, with the last four layers in MXFP8, for a ~20GB footprint. Thinking Machines ran it offline on a MacBook Pro with an M5 Pro and 48GB.
- It posts the top score on most coding and agentic search rows in the lab's comparison table, including 48.2% on DeepSWE v1.1. It trails Qwen3.8-27B on Terminal Bench v2.1 and AA-Omniscience.
- The announcement names no licence and no Tinker price, so both need checking before production use.
Inkling 2 Mini is an open-weights Mixture-of-Experts model from Thinking Machines, released on November 20, 2026. It has 26.4B total parameters and 3.4B active, with a memory footprint of about 20GB, small enough to run on consumer laptops. Thinking Machines says it leads every open-weights model under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). The announcement does not name a licence or give a price.
Inkling 2 Mini is a looped transformer: it runs its middle layers three times for every token. The release page links to the weights on Hugging Face, and the model is available on Tinker for inference and customization. Thinking Machines is also releasing a research-specific variable-loop variant, which it used for its monitorability experiments.
Looping gives 5.4B weight activations per token
The model has 24 layers in four blocks of six. The first and last blocks run once, and the two middle blocks, layers 7 through 18, run three times. Thinking Machines counts 5.4B weight activations per token to account for the repeated passes, against 3.4B active parameters. The model derives from the lab's mainline predictive scaling ladder with two changes: two-thirds of the experts were pruned to reduce size, and looping was added.
The lab reports that looping the middle 12 layers produced larger accuracy gains than looping the last 12 on 27 of 28 evaluations. It gives a compute-equivalent gain of 1.51 to 1.91x for the middle-layer position, against a range starting at 1.24x for the last layers. The speed chart's decoding figure is a theoretical ceiling at 614 GB/s with BF16 weights and batch size one, 'including repeated reads across loops'. That ceiling therefore counts the looped passes, not only the 3.4B active weights. The announcement says the model runs faster than similarly performing models, but the source text gives no measured tokens-per-second figure.
About 20GB in NVFP4 runs on a 48GB MacBook
Most of the weights were trained and are deployed natively in NVFP4, with the last four layers in MXFP8. The total footprint is about 20GB. Thinking Machines demonstrated the model offline on a MacBook Pro (M5 Pro, 48GB) in airplane mode, where it built a playable flight simulator in a few minutes. In a second demo, Thinkyville, the same model powered all 20 villagers of a pixel town on one offline MacBook.
Researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab measured intelligence per joule, the energy an LLM needs per query to reach a given level of performance. On an NVIDIA DGX Spark (128GB), Inkling 2 Mini scored 0.703. The other models scored 0.430 (Qwen3.6-35B-A3B), 0.159 (Gemma-4-26B-A4B) and 0.036 (Qwen3.5-27B). The precisions differed: Inkling ran at MXFP8 with low reasoning effort, Qwen3.6-35B-A3B and Gemma-4-26B-A4B ran at NVFP4, and Qwen3.5-27B ran at INT4.
Commenters on Hacker News focused on community runtimes. One linked the official repo at thinkingmachines/Inkling-2-Mini and an Unsloth GGUF build at unsloth/Inkling-2-Mini-GGUF. Another linked llama.cpp pull request 32695, whose description says it 'adds looped support'. A commenter running a Strix Halo Framework Desktop said they would test the model once it has llama.cpp support. The sources do not say whether the PR has merged.
One commenter also relayed a Reddit note from the team. It said Unsloth had noticed the models respond well to quantization, maintaining 90% top-1 even at 1-bit. The team said it had not exhaustively benchmarked this, and the claim does not appear in the announcement.
Inkling 2 Mini licence and pricing are unstated
The announcement says Inkling 2 Mini 'is released with open weights' but names no licence. It gives no per-token price for Tinker and no rate limits. Teams that need commercial-use terms should read the Hugging Face model card directly, because the release post does not settle them.
Ahead of Qwen3.8-27B on most coding benchmarks
The comparison table sets Inkling 2 Mini against four models. Gemma 4 31B, Muse Glimmer and Qwen3.8-27B list activated parameters equal to total (30.7B, 30.0B and 27.0B). Nemotron 3.5 Super VL has 12.0B activated and 124.5B total. Thinking Machines notes that evaluation setups vary by model. Qwen3.8-27B is the nearest competitor on most rows:
- DeepSWE v1.1: 48.2% vs 42.2% for Qwen3.8-27B.
- SWE-bench Pro v2 (Hard): 77.9% vs 73.8%.
- SWE-bench Verified, bash only: 79.2% vs 77.8%.
- SWE-bench Multilingual: 73.7% vs 70.0%.
- Terminal Bench v2.1: 71.6% vs 73.0%, one of the rows Qwen wins; Terminal Bench v4.0 is a tie at 5.6%.
- AA Index v5.0: a tie at 20.1.
- Vals Index v2.2: 46.5% vs 29.8% for Nemotron 3.5 Super VL; no Qwen score is listed.
- BrowseComp with compaction: 79.0% vs 78.9% for Nemotron 3.5 Super VL.
- PMPP-Hard kernel optimization (AUC, 0–10h): 455.82 vs 262.50 for Qwen3.8-27B; the lab says this approaches the much larger Inkling 2 Small.
Qwen3.8-27B leads on several knowledge-work evaluations: AA-Briefcase v1.1 (1391 vs 1263), GDPVal-AA v2.1 (1427 vs 1319), AutomationBench-AA v1.0.6 (48.2% vs 41.8%) and EnterpriseOps-Gym (44.2% vs 42.0%). Inkling 2 Mini leads on JobBench (49.6% vs 36.5% for Nemotron) and Vals Finance Index (50.1% vs 41.0% for Nemotron).
One Hacker News commenter called the model 'bonkers good if these benchmarks hold up in real-world scenarios'. A reply read that caveat as a sign they probably won't. Another argued that beating a 27B model 3.5 months after its release, with a new architecture at a similar parameter count, is less impressive than it sounds. A commenter planning to test on an AMD 7900XTX expected higher decode speed than Qwen3.8-27B with comparable quality, because Inkling 2 Mini is an MoE. The source text offers only the lab's qualitative speed claim and a theoretical ceiling to check that against.
Factual recall and multi-turn chat lag behind
Thinking Machines says looping improves reasoning but does not let the weights store more facts, since knowledge is mostly a function of total parameter count. On AA-Omniscience v1.1, Inkling 2 Mini scores -8.6, against -6.4 for Qwen3.8-27B and -2.9 for Nemotron 3.5 Super VL. Its non-hallucination rate is 63.4%, against 71.4% for Qwen3.8-27B.
With the variable-loop variant, extra passes improved long-context benchmarks such as GraphWalks BFS and MRCR v2. They did not improve knowledge-intensive tests such as AA-Omniscience, SimpleQA Verified, SimpleVQA and WorldVQA. The lab's answer is search and tools: Inkling 2 Mini leads open-weights models of similar size on BrowseComp, DeepSearchQA (85.3%) and MCP Atlas (82.7%).
The limitations section lists 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, and says both trail Qwen3.8-27B. The table does not match that text. Its ComplexConstraints row shows 35.6% as the top score, in the Muse Glimmer column, and 30.1% for Qwen3.8-27B, and it has no MultiChallenge row. The limitations section also says the model trails larger models on reasoning-intensive benchmarks.
On CritPt v2 Full, Inkling 2 Mini scores 31.4% against 36.9% for Nemotron, though it leads CritPt v2 Hard at 17.0% against 13.5%. On HLE-Diamond it scores 15.2%, against 15.3% for Qwen3.8-27B.
Loops cut visible reasoning, sabotage did not rise
Looping lets the model reason longer without writing its reasoning down. That affects monitoring through chain of thought. With the variable-loop variant, the no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three. Serial-reasoning performance kept improving out to ten loops, twice the number the model was trained with. The chain of thought got shorter as loops increased but stayed legible, and scores on agentic-sabotage tests did not trend higher.
On the lab's internal safety evaluations, Inkling 2 Mini posted the lowest rates in the table for misbehavior (9.8%), covert misbehavior (8.3%) and destructive actions (10.2%). These rows are labeled internal. On the public FORTRESS adversarial set it scored 89.5%, second to Muse Glimmer at 90.3%.
Questions this raises
What licence is Inkling 2 Mini released under?
The announcement says the model is released with open weights but does not name a licence. Teams that need commercial-use terms should check the Hugging Face model card directly.
Can Inkling 2 Mini run on a laptop?
Yes. Its footprint is about 20GB, and Thinking Machines demonstrated it offline on a MacBook Pro with an M5 Pro and 48GB of memory. Community GGUF builds exist, and a llama.cpp pull request adding looped support was open, but the sources do not say whether it has merged.
How does Inkling 2 Mini compare with Qwen3.8-27B?
It leads on most coding benchmarks, including DeepSWE v1.1 (48.2% vs 42.2%) and SWE-bench Pro v2 Hard (77.9% vs 73.8%). Qwen3.8-27B wins on Terminal Bench v2.1, several knowledge-work evaluations and factual recall.
Spotted an error? Email samir.s.sengupta@gmail.com and I will correct it and date the change at the top of the note. How these notes are researched, drafted and checked is in the editorial policy.
