Open sourceThinking Machines LabConfirmed
Story 1 of 13 in the Friday 20 November editionThinking Machines Lab open-sources 26B Inkling 2 Mini looped model
Apache 2.0 open weights on Hugging Face and Tinker pair a 20GB native NVFP4 and MXFP8 checkpoint with 1M context and local 32GB MacBook execution.
Thinking Machines Lab released Inkling 2 Mini (thinkingmachines/Inkling-2-Mini) on 20 November 2026 under an Apache 2.0 licence on Hugging Face alongside hosted inference and fine-tuning on Tinker, following Thinking Machines Lab’s 594B Inkling 2 Small release. Designed as a 24-layer multimodal Mixture-of-Experts looped transformer accepting text and images across a 1M-token context window, the model holds 26.4B total parameters and routes each token to 8 of 128 experts for 3.4B unique active weights. Running its two middle six-layer blocks three times with gated attention residuals yields 48 layer applications and 5.4B effective active parameters per token.
Native NVFP4 laptop deployment, agentic benchmarks and loop monitorability
Trained natively in 4-bit NVFP4 across 20 expert layers and 8-bit MXFP8 on the final four layers, the single 20GB checkpoint runs without post-training quantisation loss on Apple Silicon Macs with 32GB memory, NVIDIA DGX Spark systems or 32GB GPUs via NVIDIA Dynamo, llama.cpp, Unsloth, Ollama and Exo Labs. Mid-training pruned two-thirds of the initial 384 experts before healing on 1T distilled tokens and running 2.2T tokens of reinforcement learning.
Inkling 2 Mini scores 20.1 on Artificial Analysis Index v5.0 and leads open-weight models below 100B parameters on DeepSWE v1.1 (48.2%), Vals Index v2.2 (46.5%), SWE-bench Pro v2 Hard (77.9%) and SWE-bench Verified bash-only (79.2%). Alongside the main checkpoint, Thinking Machines Lab released TML-d24-varloop-research on Hugging Face to test chain-of-thought monitorability across one to five loop passes. Thinking Machines Lab has not announced per-token Tinker API pricing.
Technical details
- Architecture and parameter counts
- 24-layer multimodal looped MoE (12 middle layers x 3 passes = 48 steps); 26.4B total, 3.4B unique, 5.4B looped active
- Licence and checkpoint footprint
- Apache 2.0 open weights (~20GB mixed NVFP4 across 20 layers and MXFP8 on final 4 layers)
- Local hardware and runtimes
- 32GB+ Apple Silicon Macs, NVIDIA DGX Spark or 32GB+ VRAM GPUs via Dynamo, llama.cpp, Unsloth, Ollama and Exo Labs
- Context window and modalities
- Up to 1M tokens; UTF-8 text and 40px-to-4096px image inputs with text output
Release details & context
Energy efficiency on DGX Spark, CoT compression and multi-turn trade-offs
In early hardware tests by Stanford’s Hazy Research and Scaling Intelligence Labs on an NVIDIA DGX Spark (128GB) at MXFP8 precision, Inkling 2 Mini recorded 0.703 intelligence per joule, ahead of Qwen3.6-35B-A3B (0.430), Gemma-4-26B-A4B (0.159) and Qwen3.5-27B (0.036). An unpruned looped variant cost 7.5% more compute to train than a non-looped baseline yet needed 140% less inference compute to match loss on 512k-token agentic traces.
On TML-d24-varloop-research, Think Fast no-chain-of-thought 50% reliability horizons climbed from 16 seconds at one pass to 37 seconds at three passes, while GPQA Diamond reasoning traces shrank from 3,703 mean tokens at one pass to 1,661 tokens at five passes without losing English dictionary legibility, and CoT-Control-QA evasion stayed at 0.1% or lower. Thinking Machines Lab noted that Inkling 2 Mini trails Qwen3.8-27B on MultiChallenge (50.0%) and ComplexConstraints v1.1 (33.4%).
Common questions
What local hardware is required to run Inkling 2 Mini offline?
The 20GB mixed NVFP4 and MXFP8 checkpoint runs locally on Apple Silicon Macs with at least 32GB unified memory, a single NVIDIA DGX Spark system or GPUs with 32GB or more of VRAM.
Does looping middle transformer layers make Inkling 2 Mini harder to monitor via chain of thought?
Thinking Machines Lab found that five loop passes halved average GPQA Diamond reasoning length from 3,703 to 1,661 tokens by reducing backtracking, while sentence legibility held steady and deliberate chain-of-thought steering stayed at or below 0.1%.
Editor’s weekly perspective
Why it stood out: Looping middle MoE blocks three times inside a 20GB Apache 2.0 checkpoint brings 48.2% DeepSWE v1.1 coding and 1M context onto a 32GB laptop.
Released on 20 November on Hugging Face and Tinker under Apache 2.0, Inkling 2 Mini packs 26.4B total parameters (3.4B unique active, 5.4B looped active across 48 effective steps) into a roughly 20GB mixed NVFP4 and MXFP8 checkpoint running locally via llama.cpp, Ollama, Unsloth and Dynamo.
The release capped a strong 17 to 20 November open-weight run alongside Tencent’s open-weight 770B Hy4 and 220B Hy4-flash, Hugging Face’s Qwen3.8-27B-pi-adapted LoRA and MachinePulse’s Apache 2.0 Karpo Decision 31B.
Sources
- PrimaryThinking Machines Lab: Introducing Inkling 2 Minithinkingmachines.ai
- Officialthinkingmachines/Inkling-2-Mini on Hugging Facehuggingface.co/thinkingmachines
- Officialthinkingmachines/TML-d24-varloop-research on Hugging Facehuggingface.co/thinkingmachines
- On XThinking Machines Lab (@thinkymachines) Inkling 2 Mini announcement on Xx.com
- On XArtificial Analysis (@ArtificialAnlys) Inkling 2 Mini Intelligence Index thread on Xx.com
Thinking Machines Lab release timeline (3 launches tracked)
- Thinking Machines Lab debuts 594B Inkling 2 Small MoE on Tinker with Apache 2.0 weights planned
Thinking Machines Lab launched Inkling 2 Small (594B total, 17B active) on Tinker, scoring 31 on the AA Intelligence Index with Apache 2.0 weights planned.
- Thinking Machines Lab releases Inkling 2 Small weights on Apache 2.0
Thinking Machines Lab releases open weights for its 594B-total, 17B-active Inkling 2 Small multimodal MoE on Hugging Face under an Apache 2.0 licence.

