
Release Scope and Open-Weight Positioning
Thinking Machines Lab, Inc. has published thinkingmachines/Inkling-2-Mini on Hugging Face, distributing an open-weight multimodal model under the Apache 2.0 license. Categorized for image-text-to-text generation and conversational workloads, the release pairs an approximately 20GB checkpoint in mixed NVFP4 and MXFP8 precisions with support for a context window of up to 1M tokens. While the Hugging Face listing carries a publication timestamp of November 17, 2026 UTC, the model card itself states a release date of November 20, 2026. Intended for English, broader multilingual tasks, and multiple programming languages, Inkling 2 Mini targets developers building agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation pipelines. Alongside direct weight downloads, Thinking Machines Lab provides managed API and fine-tuning access through its Tinker platform, supported by the Tinker Cookbook and the tml-renderers package, as well as inference-only access in the Tinker Playground.
Looped MoE Architecture and Reported Benchmarks
Under the hood, Inkling 2 Mini implements a multimodal autoregressive looped transformer, designated in repository metadata as inkling2_loop. Rather than scaling parameter storage linearly with effective depth, the 24-layer decoder-only transformer uses sparse Mixture-of-Experts (MoE) feed-forward layers and executes its middle 12 layers three times with shared weights. That recurrence delivers 48 layer applications per token from 24 stored layers. Within each layer, every token routes to 8 of 128 experts, paired with a hybrid of local and global attention layers. A vision transformer encodes pixel-based images—with optimal dimensions specified between 40px and 4096px—and projects them alongside UTF-8 text into a shared hidden space processed jointly by the decoder. Total capacity is 26.4B parameters with 3.4B base active parameters; because each additional pass through the looped block adds roughly 1.0B effective active parameters, three passes yield 5.4B effective active parameters per token.
In self-reported evaluations published on the model card, Thinking Machines Lab reports competitive coding and agentic results against larger peers, though independent verification is not included in the primary listing. On SWEBench Verified using bash only, the company reports 79.2%, compared with 77.8% for Qwen3.8-27B and 71.1% for Nemotron 3.5 Super VL. On SWEBench Pro v2 (Hard) with mini-swe-agent v2.4.6, it lists 77.9%, alongside 73.7% on SWEBench Multilingual with bash only, 71.6% on Terminal Bench v2.1 using pi v0.83.0, 48.2% on DeepSWE v1.1, 15.1% on FrontierCode v2.0 with stirrup v0.2.1, and 5.6% on Terminal Bench v4.0. On aggregate indices, the provider reports 20.1 on AA Index (v5.0) and 46.5% on Vals Index (v2.2).
Local Hardware Targets and Inference Integration
Shipping a single 20GB checkpoint in mixed NVFP4 and MXFP8 formats narrows the hardware barrier for local multimodal and agentic experimentation. Thinking Machines Lab lists verified support for a single NVIDIA DGX Spark as well as Apple silicon devices—including MacBooks, Mac minis, and Mac Studios—configured with 32GB or more unified memory. Other GPUs and platforms are also supported, with at least 32GB of VRAM recommended. Software ecosystem coverage at launch spans major open-source serving stacks. Developers can run the model through Hugging Face transformers using AutoModelForSeq2SeqLM or the image-text-to-text pipeline, or serve OpenAI-compatible endpoints via vLLM and SGLang, including official SGLang Docker images. Containerized local execution is also documented for Docker Model Runner via docker model run. Beyond primary GPU servers, the card notes support for TokenSpeed, NVIDIA Dynamo, Unsloth, and llama.cpp for quantized GGUF builds, plus Llama App, Ollama, and Exo Labs on Apple silicon. For post-training workflows, the weights support Miles and SkyRL.
Licensing, Deployment, and Verification Caveats
Despite the permissive Apache 2.0 weight license, engineers should weigh several operational and governance constraints noted in the documentation. First, Thinking Machines Lab pairs the Apache 2.0 grant with a separate Model Acceptable Use Policy referenced in the model card, which downstream integrators should review before commercial rollout. Second, while the 20GB checkpoint fits comfortably within 32GB memory envelopes, exercising the full 1M-token context window or processing high-resolution images up to 4096px will increase KV-cache and activation memory demands well beyond the base file footprint. Third, training data disclosures cover text, images, and video from public internet sources, public repositories, third-party acquisitions, internal generation, and synthetic augmentation with deduplication and filtering, without publishing dataset-level manifests. Finally, all benchmark comparisons against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL span varying agent harnesses and reflect company-reported numbers rather than third-party audits, making local validation essential prior to production deployment.
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-21.



