deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

Thinking Machines releases Inkling 2 Mini, a looped-transformer model built for laptops

Thinking Machines has released Inkling 2 Mini, a 26.4B-parameter open-weights looped transformer that runs in about 20GB of memory, claiming leading coding results among open models under 100B parameters.

Thinking Machines releases Inkling 2 Mini, a looped-transformer model built for laptops

What Thinking Machines released

Thinking Machines has released Inkling 2 Mini, an open-weights model that the company says is small enough to run offline on consumer laptops and desktop AI machines. The release, which surfaced on Hacker News's front page, is built around an unusual looped-transformer architecture, and the company also published a research variant with adjustable loop depth for studying that design.

According to the announcement, Inkling 2 Mini is a Mixture-of-Experts model with 26.4 billion total parameters, of which 3.4 billion are active per token. Because its middle layers execute three times for each token, the effective weight activation per token rises to 5.4 billion. The company says the model also decodes faster than similarly performing alternatives.

How the looped design works

Rather than growing the parameter count, looping reuses capacity: the middle of the network runs repeatedly, giving the model extra rounds of computation on every token. Thinking Machines says this allows deeper reasoning per token and correct answers with fewer tokens, which it frames as especially valuable for reasoning-heavy work such as agentic coding.

Looping also has a safety dimension the company highlights. It lets a model reason internally without writing all of that reasoning into its chain of thought, which affects monitorability — the ability to understand what a model is doing by reading its output.

Measurements with the variable-loop variant showed a no-chain-of-thought time horizon improving from 16 seconds at one pass to 37 seconds at three, and performance on a serial-reasoning benchmark kept improving out to ten loops — twice as many as the model was trained with. The chain of thought got shorter as loops were added but stayed legible, and scores on tests of agentic sabotage did not trend upward with more loops.

Benchmark results

Thinking Machines claims the model tops all open-weights releases below 100 billion total parameters on benchmarks including DeepSWE v1.1, at 48.2%, and Vals Index v2.2, at 46.5%. It also reports 77.9% on SWE-bench Pro v2 (Hard), 79.2% on SWE-bench Verified with bash only, and 79.0% on BrowseComp using context compaction.

Comparisons in the announcement cover Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. Inkling 2 Mini leads most coding, search and tool-use rows, but not all: Qwen3.8-27B is ahead on Terminal Bench v2.1 (73.0% versus 71.6%) and MMMU Pro (76.3% versus 74.9%), while Nemotron 3.5 Super VL leads on AA-Briefcase (1391 versus 1263) and GDPVal-AA (1427 versus 1319). The company positions the model as strongest on long-horizon agentic work, mirroring its larger sibling Inkling 2 Small, and says it approaches that larger model on PMPP-Hard, a kernel-optimization benchmark.

Where looping helps and where it doesn't

Looping buys reasoning, not knowledge. Thinking Machines notes that stored knowledge is mostly a function of total parameter count, and looping doesn't change that: on AA-Omniscience, a factuality measure, the model lands in line with same-size peers. In the variable-loop experiments, long-context tests such as GraphWalks BFS and MRCR v2 improved with more passes, while knowledge-heavy tests including SimpleQA Verified, SimpleVQA and WorldVQA did not.

The model instead compensates through search and tools, leading comparable open-weights models on BrowseComp, DeepSearchQA and MCP Atlas.

Built for local hardware

Most weights were trained and deployed natively in NVFP4, a 4-bit floating-point format, with the final four layers in MXFP8. The total memory footprint is roughly 20GB, which is what makes laptop-scale deployment practical. As a demonstration, researchers ran the model offline in airplane mode on a MacBook Pro with an M5 Pro chip and 48GB of memory, where it built a playable flight simulator within a few minutes.

The weights are open, and the model is available on Tinker for inference and customization.

Why it matters

Looping is a different axis of scaling: instead of paying for capability with raw parameter count, Inkling 2 Mini trades compute per token for a smaller memory footprint. For local inference, where RAM is the binding constraint, that is a meaningful trade-off, putting strong agentic-coding performance inside roughly 20GB. The monitorability findings matter too, suggesting a model can reason more internally without its chain of thought becoming unreadable or more evasive — though these were measured by Thinking Machines on its own variants. By open-sourcing both the fixed-loop model and the variable-loop research model, the company is giving outside researchers an unusual opportunity to probe the architecture directly.

  • #open-weights
  • #local-inference
  • #language-models
  • #mixture-of-experts
  • #quantization

Related posts