Daily Frontier Digest ·
Inkling 2 Mini: a 26B open model that thinks by looping its middle layers
Curated by Alfredo Borrás · Plymouth AI Lab
Thinking Machines releases Inkling 2 Mini, an open model that repeats its middle layers three times per token; METR says public evidence can no longer rule out fast AI R&D acceleration; OpenAI reports extraction attacks on GPT-6.2 Astra's hidden reasoning down to 0.2%; Levent Alpöge claims a Lean-formalised proof of the Riemann Hypothesis made with Claude; and a 625-person trial links two weeks off ChatGPT to less loneliness.
The story today
Inkling 2 Mini: a 26B open model that thinks by looping its middle layers
Thinking Machines Lab released Inkling 2 Mini on 20 November under the Apache-2.0 licence. It is a mixture-of-experts model with 26.4B total parameters and 3.4B active per token. Its 24 layers are split into four blocks of six; the first and last blocks run once, while the two middle blocks run three times for every token, so each token gets roughly twice the depth the weights alone would suggest. Most weights were trained and ship in NVFP4, which puts the whole model at about 20 GB, and the lab says it runs through Ollama on any Mac with at least 32 GB of RAM.
On the lab's own comparison table it scores 20.1 on the Artificial Analysis Index v5.0, tied with Qwen3.8-27B, and 48.2% on DeepSWE v1.1, which the lab says leads all open-weight models under 100B parameters. The lab also notes that evaluation setups differ between models, and it lists weak spots: 50.0% on MultiChallenge and 33.4% on ComplexConstraints, both behind Qwen3.8-27B.
The interesting part is what looping does and does not buy. In a research variant with an adjustable number of passes, long-context scores rose with more loops but knowledge benchmarks such as SimpleQA Verified did not move: extra passes add computation, not stored facts. Because some of that reasoning now happens inside the loops rather than in written text, the lab also checked monitorability, reporting that the chain of thought got shorter with more loops but stayed readable, and that sabotage test scores did not rise.
Why it matters: If you want a capable local coding agent on a 32 GB machine, Inkling 2 Mini is worth a test run, but pair it with search or retrieval for factual work, and run your own benchmarks: the headline coding numbers are the maker's.
Sources: Thinking Machines Lab: Introducing Inkling 2 Mini · Hugging Face: thinkingmachines/Inkling-2-Mini
Highlights
METR: current evidence cannot rule out drastic AI R&D acceleration within a couple of years
In its Q4 2026 capability update, covering public models up to Opus 5.5 and GPT-6 Astra, METR says public evidence is no longer enough to exclude a very fast speed-up of AI research in the next couple of years, while stressing large uncertainty. On its combined task suites, the length of task models can finish half the time is doubling about every 3.6 months, and frontier models reach an estimated 27 hours on the verifiable part of its newer, harder suite. The caveats are METR's own: actual human attempts underpin only 9% of its human-time estimates, the rest are estimated, and models do worse on tasks that are costly to verify and need judgment, which METR names as a likely brake.
safetyevalsSource: METR: Q4 2026 capability update
OpenAI says attacks on GPT-6.2 Astra's hidden reasoning now succeed 0.2% of the time
OpenAI described how it fights adversarial distillation, where someone trains a copycat model on its outputs. In an internal automated red-team test, attempts to recover reasoning a model produced for an earlier prompt succeeded 54.0% of the time against GPT-6 Astra with its old safeguards and 0.2% against GPT-6.2 Astra; for the current prompt the figures were 22.9% and 0.9%. Defences include adversarial training, input and output classifiers that also scan tool-call arguments, and tying reasoning to the conversation that produced it. An internal study found a student model trained on exact reasoning traces generally picked up more cyber skill than one trained on summaries. All results are OpenAI's own and the student model is not named.
safetydistillationSource: OpenAI: Combatting adversarial distillation
Levent Alpöge claims a Lean-checked proof of the Riemann Hypothesis, found with Claude
Mathematician Levent Alpöge posted a four-page note saying that, working with Claude, he believes he has proved the Riemann Hypothesis and that the method also gives the Generalized Riemann Hypothesis. Instead of a readable paper he released a Lean formalisation split into 14 archive parts, saying a full written version would be an enormous, unreadable document (Claude's draft passed 250,000 pages). His timeline runs from 6 October to 16 November 2026. Treat this as a claim: no independent mathematician has yet confirmed that the formal statement proved matches the hypothesis as mathematicians mean it, and the author himself says he has not yet mastered all of the mathematics involved.
mathsleanSource: Levent Alpöge: remark (PDF)
Cursor tells House committees it may phase out Chinese open-weight models
CNBC reports that a House report from the Homeland Security Committee and the Select Committee on China, published on 20 November, says Cursor informed the committees it will explore phasing out China-developed open-weight models because of geopolitical risk. Cursor, which SpaceX finished acquiring in August, built its Composer 2 coding model on Moonshot AI's Kimi K2.5. The report credits cost, customisation and near-frontier performance for the spread of Chinese open models and recommends a national American open-weight initiative and barring PRC-developed models from federal systems and contracts. No timeline or model list is given, and Cursor did not respond to CNBC.
industrypolicySource: CNBC: SpaceX's subsidiary Cursor tells lawmakers it's considering phasing out Chinese AI models
Signal will block OS-level AI assistants from capturing its screen
Signal president Meredith Whittaker wrote that Signal will block screenshots in the app on iOS 27 and block screen capture on Android versions where it can detect an AI assistant trying to read the screen, citing Gemini Intelligence and Siri's Onscreen Awareness. It is the same approach Signal took against Windows Recall last year. Users can turn the protection off in settings. Signal calls its fix incomplete because agents could still read app data already on the device through the filesystem, and says it cannot yet do this on macOS because operating-system makers give developers no clean way to restrict AI access.
industryagent-securitySource: Signal: Signal is Serious About Protecting Your Privacy
Concept of the day
Looped transformer
A looped transformer reuses some of its layers several times for each token instead of stacking more distinct layers. Inkling 2 Mini runs its first six layers once, its middle twelve layers three times, and its last six once, so 24 physical layers do the work of 48. The repeated passes give the model more computation per token without adding weights or memory, which helps reasoning, but they cannot add knowledge the weights never learned, and any thinking done inside the loops is not visible as written chain of thought.
Models released
| Model | Maker | Size | Notes |
|---|---|---|---|
| Inkling 2 Mini | Thinking Machines Lab | 26.4B total, 3.4B active (5.4B weight activations per token with looping); about 20 GB | Apache-2.0 open weights. Looped mixture-of-experts: middle 12 of 24 layers run three times per token. Trained and shipped mostly in NVFP4 with the last four layers in MXFP8; runs via Ollama on Macs with 32 GB+ RAM, GGUFs from Unsloth. A variable-loop research variant is released alongside it. |
Try this
Train a small LoRA for the agent harness you actually use
Hugging Face trained one LoRA adapter per agent harness (pi, Hermes Agent and OpenCode) for a frozen Qwen3.8-27B, using reinforcement learning inside the real harness with TRL's AsyncGRPOTrainer. In pi, Terminal-Bench 2.1 pass@1 rose from 39% to 55% while average tokens per task fell from 2.9K to 1.5K, mostly because the model reasoned less and made fewer tool calls. The three adapters were also concatenated into one merged adapter that runs in all three harnesses; at half weight it scored 52.5% averaged across them against 43.3% for the base model. The recipe, tasks, adapters and GGUFs are published, so you can repeat it for your own harness; the scores are Hugging Face's own.
harnesseslocaltry-thisSource: Hugging Face: Harness adapters at scale
Research
Two weeks without AI assistants reduced loneliness in a 625-person trial
Researchers including Dora Zhao, Robert Kraut and Diyi Yang randomly assigned 625 active ChatGPT users to stop using AI assistants for 14 days or for one day. Most used AI for work: 84.7% said productivity was their main use. The two-week group reported lower loneliness than the control group (β = -0.10, p = 0.03), and the gap remained two weeks after they went back to AI (β = -0.16, p = 0.02). They also met more people in person and spent more time socialising, often asking others for help they would have asked AI for. The effects are small and loneliness is self-reported; it is a preprint.
researchwellbeingSource: arXiv: AI Deactivation Reduces Loneliness and Increases Social Interaction
Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.
Related service: AI automation services for logistics
Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free