Thinking Machines released Inkling 2 Mini on November 20, 2026, and vLLM added day-0 serving support for both it and the TML-d24-varloop-research checkpoint. According to vLLM, Inkling 2 Mini is a multimodal Mixture-of-Experts model with 26B total parameters, 3.5B active parameters and up to 1M tokens of context. Its middle 12 layers run three passes per token with shared weights, giving each pass its own KV cache and gated messages from the previous pass. For reinforcement learning, vLLM also serves the model with SkyRL and reports that its MoE outputs match the reference implementation bitwise given the same inputs.13

How vLLM runs the looped layers

vLLM registers each extra loop pass as its own attention layer, letting existing prefix caching manage all 48 KV caches and 48 conv-state caches. On TML-d24-varloop-research, builders can set loop_num_passes or loop_share_cache via --hf-overrides, though both options are rejected on Inkling 2 Mini and DSpark requires the default three passes. On one GB200 at TP1 with 4K context, vLLM reports that one pass cuts batch-1 decode latency to 1.51 ms, down from 3.85 ms at three passes. That change raises KV cache capacity from 175K tokens to 352K tokens.1

How it works

Where to get it and what it needs

According to the Hugging Face model card, Inkling 2 Mini uses the Apache 2.0 license and ships as a single 20GB checkpoint in mixed NVFP4 and MXFP8 precisions. Developers can download the weights from Hugging Face or use the Tinker platform for API access and inference. The model card recommends 1x NVIDIA DGX Spark, Apple silicon Macs with at least 32GB of memory, or other GPUs with at least 32GB of VRAM.2

What changed

How it measures up on coding and efficiency

Thinking Machines reports that Inkling 2 Mini targets agentic coding and tool-use workflows. On SWEBench Pro v2 (Hard) with mini-swe-agent v2.4.6, Thinking Machines reports a 77.9% score for Inkling 2 Mini, compared with 73.8% for Qwen3.8-27B. Researchers at Stanford's Hazy Research Lab and Scaling Intelligence Lab also measured energy efficiency on an NVIDIA DGX Spark (128GB). Running Inkling 2 Mini at MXFP8 with low reasoning effort, they reported an intelligence-per-joule score of 0.703, compared with 0.430 for Qwen3.6-35B-A3B at NVFP4.3

How it works

Stated limits for builders

The model card states that Inkling 2 Mini trails Inkling 2 Small across most evaluations, especially on factual recall and knowledge-heavy questions. In internal testing, the model sometimes entered repetitive reasoning loops while debugging code, so Thinking Machines recommends setting output-token and agent-turn limits. It also advises running agentic workloads in a sandbox because the model may take destructive actions without confirmation.2