Twenty gigabytes. That is all the memory Thinking Machines' new Inkling 2 Mini needs, and in that footprint it scores 79.2% on SWE-bench Verified. I spent Saturday reading the details instead of the latest Nvidia boardroom gossip, and I think this is the most consequential release of the month for people who actually ship code. Not because it beats the frontier. Because it doesn't need a building.
Inkling 2 Mini gets its depth by looping, not by growing#
VentureBeat's Carl Franzen has the full breakdown. It is a sparse mixture-of-experts with 26.4B total parameters across 24 layers, routing each token to 8 of 128 experts, so about 3.4B parameters fire per pass. The clever part is the loop: the first six and last six layers run once, while the middle 12 run three times with shared weights. That yields 48 effective layer applications and 5.4B effective active parameters per token without storing a single extra weight.
Thinking Machines says a non-looping baseline needed 140% more inference compute to match the looped model's loss on 512K-token agentic traces. In-house comparisons deserve some salt, but the third-party numbers point the same way:
79.2% on SWE-bench Verified and 77.9% on SWE-bench Pro v2 Hard
$0.59 per task versus $3.53 for Qwen3.8-27B, and 9.0 minutes versus 38.2, according to Artificial Analysis
0.703 on Stanford's intelligence-per-joule metric, against 0.430 for Qwen3.6-35B-A3B
$0.15 / $0.50 per million input/output tokens on the first-party API, $0.08 / $0.35 on OpenRouter
The weights are Apache 2.0 on Hugging Face, with Ollama and EXO support and GGUF builds from Unsloth that bring it down to 16GB machines. Early reports put decode at 125 tokens per second on a 48GB M5 Max and 104 on an Nvidia 5090. The model takes text and images and offers a 1M-token context window.
It has real gaps, though. It trails Qwen3.8-27B on MultiChallenge (50.0%) and ComplexConstraints (33.4%), its non-hallucination rate on AA-Omniscience is 63.4% against Qwen's 71.4%, and VentureBeat notes it can fall into repetitive debugging loops. That last one is the weakness I'd watch. A local agent that spins in circles costs you nothing in tokens and everything in patience.
The interpretability work is the part I didn't expect. Thinking Machines shipped a companion variable-loop checkpoint trained for one to five passes, and NDIF and Goodfire are studying it. The company reports the no-chain-of-thought reasoning horizon rising from 16 seconds at one pass to 37 at three, while chain-of-thought length fell by more than half. Silent reasoning in the loop is efficient, and it is also exactly the kind of thinking you can't read in a transcript. I'm glad someone is poking at it from day one.
My plan for this week: pull the GGUF, point a local agent harness at a real repo, and run it against the routine refactors I currently pay a hosted model for. If the per-task numbers hold even halfway, the default for boring tickets changes.
Meta's AI buildout just outran its climate promise#
While compute shrank onto a laptop, the other end got heavier. Ars Technica reports that Meta's 2026 sustainability report, dated November 19, says the company is re-evaluating its 2030 net-zero goal after a record percentage increase in greenhouse gas emissions, one Ars says exceeds any other Big Tech competitor building AI data centers.
“Our long-term ambitions haven’t changed, but our operating environment has.” — Rachel Peterson and Blair Swedeen, Meta 2026 sustainability report
The stated reason: “demand for AI infrastructure has grown faster than almost anyone projected.” The 2030 timeframe and the near-term intensity targets are both on the table. I read that as the deadline going away while the ambition stays on the slide. Ars asks whether others will follow, and I'd bet on yes.
A year ago intelligence-per-joule looked like an academic curiosity. Next to Meta's report it reads like a line item.
Trump wants server farms at home, even where towns are saying no#
Then Washington weighed in. At a press conference on Friday, Trump promised a “tremendous tax” on companies that build data centers abroad, the Verge's Lauren Feiner reports. “We’re not going to allow our companies to build data centers outside of our country,” he said, adding that he hadn't even told his own staff yet.
The squeeze is plain. States and cities are passing moratoriums and new restrictions on data center applications, and the threat is meant to make going abroad less attractive as a way out. Feiner points out that a Democratic-majority Congress will be sworn in within a couple of months, so turning this into law is a long shot. Still, if you plan capacity, the direction is what counts: building at home keeps getting harder, and leaving might get taxed.
Put the Meta and Trump stories side by side and the hyperscalers' bind is obvious. Every new gigawatt needs a town that agrees and a climate goal that bends. A 20GB model on hardware people already own needs neither.
Also today#
UN talks in Geneva on autonomous weapons ended with countries agreeing only to “start discussing options,” with two years to propose rules, some of which may stay optional. NPR named Russia, Israel and the US as the main obstacles; the US proposed a five-year window for nonbinding guidelines. The Verge
Microdrama startup Shortical premiered Inevitable, a fully AI-generated series starring actor Aki Avni as a licensed digital avatar of himself. Its “License Your Face” program hasn't signed anyone else yet. TechCrunch
Small looped models won't take every job, and the benchmarks above show exactly where Inkling slips. But each time a laptop model clears a bar like 79% on SWE-bench Verified, another slice of everyday work stops needing a $10 billion campus, while the politics around those campuses keep getting worse. My prediction: by spring, the interesting question won't be whose frontier model is best. It will be which tasks still justify a round trip to somebody else's data center.