TML-d24-varloop: Thinking Machines' research model with a variable loop count
Thinking Machines Lab released TML-d24-varloop-research, a 26.4B-parameter open-weights research model whose middle 12 layers can be looped 1-5 times at inference, letting researchers dial effective active parameters from 3.4B to 7.4B per token on one set of weights.
Why it matters
The first public model that exposes loop count as an inference-time knob, enabling clean controlled experiments on how extra looped computation changes a model's internal computation, on a single checkpoint.
What you could build with it
A researcher could use the variable loop count to run controlled experiments on test-time compute, measuring exactly how each extra pass through the shared layers changes reasoning quality on the same weights.
Does it hold up?
Too early to judge: a research artifact released Nov 20; the lab itself warns it may behave unpredictably at loop counts other than 3 and has not evaluated it in depth.
Built with TML-d24-varloop
- Welcome Inkling 2 Mini: Cutting-edge Performance for your Local Setuphuggingface · Hugging Face's walkthrough of Inkling 2 Mini and the variable-loop research variant, with transformers, llama.cpp, SGLang and MLX inference code including how to set loop_num_passes.
- TML-d24-varloop-research-GGUF: GGUF quantization of the variable-loop modellocalmodelwatch · Local Model Watch listing of the GGUF quantized release of TML-d24-varloop-research for local llama.cpp inference.
- ggml-org/TML-d24-varloop-research-GGUFgithub · Official ggml-org GGUF quants of the variable-loop research model (IQ4_NL and Q8_0), referenced in the Hugging Face blog post.
Learn more
First spotted on article: source.
More AI models
MiniMind: train your own 64M LLM from scratch in about 2 hoursA complete, from-scratch LLM training pipeline at 64M parameters - pretraining, SFT, LoRA, DPO, PPO/GRPO, tool use and…models · JEV 0.82VoiceStudio: the open-source, fully-local ElevenLabs alternativeA desktop app bundling 20 local speech-generation engines and 10 transcription options behind one UI — voice cloning…models · JEV 0.76Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73tiny-voca: a fully open 4.5M-parameter TTS trained for under $10 of computeA fully open 4.5M-parameter English TTS — code, weights and dataset under Apache 2.0 — trained from scratch in 7.7…models · JEV 0.67Qwen4-Max: Alibaba's new flagship model goes liveAlibaba's 2.4T-parameter flagship is now generally available via API, with a 1M-token context, native vision, and…models · JEV 0.67
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.