Thinking Machines Open-Sources Inkling 2 Mini, a 26B Looped Model for Local Agents
On this page 8 sections
TL;DR
Thinking Machines Lab open-sourced Inkling 2 Mini on November 20: a 26.4B-parameter multimodal MoE that runs its middle 12 layers three times per token, so 3.4B active parameters per pass do the work of roughly 5.4B while the checkpoint stays near 20GB. Weights are on Hugging Face under Apache 2.0, it runs on a 32GB laptop, and company-run benchmarks put it at 48.2% on DeepSWE v1.1, which the lab says leads every open-weights model under 100B total parameters. A variable-loop research model shipped the same day so outsiders can test whether looped reasoning stays readable.
What shipped
Inkling 2 Mini is a 24-layer decoder with sparse MoE feed-forward layers, text and image input, and a 1M-token context window. The checkpoint is about 20GB, trained and deployed in NVFP4 with the last four layers in MXFP8, so there is no quantization conversion loss between training and inference. Supported hardware on day one: Apple silicon with 32GB or more of memory, one NVIDIA DGX Spark, or other GPUs where 32GB or more of VRAM is recommended.
The software support is unusually broad for a two-day-old architecture: SGLang and vLLM recipes, Unsloth GGUF builds, llama.cpp support, and Ollama on Apple silicon, plus hosted fine-tuning through Tinker. Miles and SkyRL post-training recipes already exist for it.
Both launch demos ran offline. On a MacBook Pro with an M5 Pro and 48GB, in airplane mode, the model built a playable flight simulator in a few minutes. A second demo runs 20 simulated villagers as separate agents from one offline MacBook. A model writing a flight simulator mid-flight, with its Wi-Fi off, is the right amount of showing off.
The loop is the product
The 24 layers split into four blocks of six. The first and last blocks run once each; the two middle blocks, layers 7 through 18, run three times with shared weights. That gives 48 layer applications per token from 24 stored layers, with each token routed to 8 of 128 experts per layer. Passes hand a "message" forward through a learned gate that starts closed, so training begins like any other model, and opens as training progresses. The model also uses attention residuals, where each layer reads a learned mix of earlier blocks' outputs rather than a single running sum.
A standard transformer writes each intermediate thought down once and passes it up the stack. Inkling 2 Mini rereads the middle of its notes three times before answering, instead of buying a bigger notebook.
The payoff, from Thinking Machines' own scaling runs: at training step 2,000, a no-loop baseline needed an estimated 1.51 to 1.91 times as much compute to match the middle-loop model's loss, against 1.24 to 1.36 times for a last-layers loop. A non-looping sibling cost 7.5% less to train and still needed 140% more inference compute to match the looped model's loss on 512k-token agentic traces. The idea goes back to Universal Transformers in 2018; this is one of the first production open-weights versions of it.
The recipe for getting small is published too. Training started with 8 experts chosen out of 384; mid-training, two-thirds of the experts were pruned away, leaving the 128 most used per layer. A healing phase of 1T tokens distilled from a larger internal Inkling model recovered the lost performance, followed by SFT and RL totaling 2.2T tokens over 2,000 steps.
The numbers, and their asterisk
Company-run agentic coding results first, with the caveat that evaluation setups vary by model and no independent leaderboard has the model yet:
Elsewhere in the vendor table: 79.2% on SWE-bench Verified with bash only, ahead of Qwen3.8-27B at 77.8%, plus 77.9% on SWE-bench Pro v2 Hard, 46.5% on Vals Index v2.2, and a PMPP-Hard kernel-optimization AUC of 455.82 against 262.50 for Qwen3.8-27B. Stanford's Hazy Research and Scaling Intelligence labs, which had early access, measured intelligence per watt at 0.703 on a DGX Spark, against 0.430 for Qwen3.6-35B-A3B and 0.159 for Gemma 4 26B A4B.
The lab's own limitations section does honest work. MultiChallenge lands at 50.0% and ComplexConstraints v1.1 at 33.4%, both behind Qwen3.8-27B, so multi-turn chat and instruction-heavy prompts with many constraints are weak spots. Knowledge tracks total parameter count and looping adds none of it; AA-Omniscience scores -8.6 against Qwen3.8-27B's -6.4. Where the model needs facts it leans on search and tools, at 79.0% on BrowseComp with compaction. On the vision side it takes OmniDocBench v1.6 at 92.8% and ScreenSpot Pro at 76.5% raw, 87.7% with Python.
Why safety people watched this one
Looping moves computation off the page. A chain-of-thought monitor can only read what the model writes down, and a looped model can think harder per token before writing anything, which cuts at the assumptions CoT monitoring rests on. Thinking Machines says it released the model partly so researchers can poke at exactly this, and shipped TML-d24-varloop-research, a twin trained at one to five passes, to make depth comparisons clean.
The preliminary findings: on Think Fast, the variable-loop model's no-CoT time horizon rose from 16 seconds at one pass to 37 at three, and on N-Hop Lookup, a pure serial-reasoning test, accuracy kept improving out to ten passes, double the maximum depth it trained with. More loops also compressed the written reasoning. On GPQA Diamond questions answered correctly at every depth, five-pass reasoning averaged 1,661 tokens against 3,703 at one pass, while sentence length and the share of dictionary words stayed flat. The visible difference was less backtracking: uncertainty markers like "wait" and "actually" fell from 9.7 to 3.6 per answer. Scores on agentic sabotage tests did not trend up with loop count.
One evaluation came out uninformative. Controllability of the chain of thought was poor for every model measured, with Mini following such instructions at most 0.1% of the time against 1.3% for Qwen3.8-27B, too low a rate to separate loop depths. At this scale the lab found no meaningful drop in monitorability, and it published the setup so anyone can recheck.
Day one in the field
The Hacker News thread reached 388 points, and the Hugging Face repo logged about 7.9K downloads and 3.2K likes in two days. Field reports skew positive on coding: a 3090 owner who ran it all day calls it a workhorse, a laptop RTX 4060 with 8GB reports 6.7 tok/s, and a Strix Halo setup with optimized kernels and speculative decoding reports 30 to 45 tok/s. AMD users initially saw roughly a third of Qwen-class speed because the engines had not caught up, after which third-party runtimes pushed past 100 tok/s. A new architecture arriving faster than its inference engines is a tradition at this point.
If you build local agents
- The floor is 32GB of unified memory or VRAM, and 20GB of weights leaves a 24GB card little headroom once you cache context. Budget the KV cache before you budget the model.
- Memory math and compute math differ: you store 26.4B parameters but pay roughly 5.4B effective active per token, and only on engines that implement the looped MoE well. Benchmark in your own harness before promising throughput.
- Match jobs to the strengths: agentic coding, search-and-tool work, long-context retrieval, document vision. Keep a chattier model on conversational, constraint-heavy surfaces.
- Fine-tuning paths exist already: Tinker hosted, plus Miles and SkyRL recipes.
- The architecture is copyable and the monitorability question now has a public testbed. More looped models will follow; the varloop model card is where to start reading.
Key Takeaways
- Inkling 2 Mini shipped November 20 under Apache 2.0: 26.4B total parameters, 3.4B active per pass, roughly 5.4B effective active per token because the middle 12 of 24 layers run three times, in a ~20GB native NVFP4/MXFP8 checkpoint.
- It targets 32GB+ Apple silicon, one DGX Spark, or 32GB+ GPUs, with SGLang, vLLM, Unsloth GGUF, llama.cpp, and Ollama support at launch and fine-tuning via Tinker.
- Company-run benchmarks: 48.2% DeepSWE v1.1 and 46.5% Vals Index v2.2, a claimed lead over every open-weights model under 100B, plus 79.2% SWE-bench Verified bash-only. Independent numbers are not out yet.
- Weak spots per the lab's own table: multi-turn chat (MultiChallenge 50.0%) and constraint following (33.4%). Looping buys serial reasoning, and knowledge stays capped by parameter count.
- The variable-loop research twin shows no-CoT capability rising with passes (16s to 37s time horizon) while written reasoning shrinks by half and stays legible; sabotage metrics stayed flat and the controllability test was too weak to separate models.
- Builder math: store 26.4B, pay ~5.4B active per token, and plan KV cache against the 20GB checkpoint on smaller cards.
Sources: Thinking Machines Lab, "Introducing Inkling 2 Mini" (Nov 20, 2026), Inkling 2 Mini model card, Hugging Face, TML-d24-varloop-research model card, Hacker News discussion