Skip to content
AIBRAINWORX / THE INDEPENDENT AI LABRESEARCH · HARDWARE · OPEN INTELLIGENCE
Models & Weights / Inference & Hardware / Model Watch

Thinking Machines Lab Releases Inkling 2 Mini Looped Multimodal MoE for Local Inference

Thinking Machines Lab introduces Inkling 2 Mini, a 26B-total, 3B-active multimodal Mixture-of-Experts model using a three-pass looped decoder and 1M-token context for local and datacenter deployment.

Listen to this articleBrowser Voice
Ready
Official company logo or product image for thinking-machines-lab-inkling-2-mini-looped-multimodal-moe
Official image or organization logo from huggingface.co; identification only, no endorsement.
Event / context: 2026-11-20. Source check: October 28, 2026.

Thinking Machines Lab has released Inkling 2 Mini, a compact multimodal Mixture-of-Experts (MoE) language model engineered to bring higher effective parameter capacity to local hardware setups without expanding checkpoint storage. Detailed in a Hugging Face technical blog post published November 20, 2026, by Merve, Pedro Cuenca, Aritra Roy Gosthipaty, and Guarin Flück, the release pairs a 26-billion-parameter physical footprint with 3 billion active parameters per token while supporting both text and image inputs across a 1-million-token context window.

Core Announcement and Local-First Scope

Positioned as a substantially smaller sibling to Inkling 2 Small, Inkling 2 Mini (thinkingmachines/Inkling-2-Mini) targets developers and practitioners running inference on local workstations alongside production datacenter deployments. Thinking Machines Lab ships the flagship weights in a mixed-precision format combining NVFP4 and MXFP8 expert weights, paired with a DSpark drafter designed to accelerate speculative decoding at a modest memory overhead. Alongside the fixed-loop production checkpoint, the lab also published an experimental research artifact, thinkingmachines/TML-d24-varloop-research, which exposes configurable loop depth during inference.

Looped Transformer and MoE Architecture

Rather than stacking 48 distinct physical decoder layers, Inkling 2 Mini uses 24 physical layers and executes its middle core in a three-pass loop. Layers 0 through 5 run once, layers 6 through 17 repeat three times using shared weights, and layers 18 through 23 run once, yielding 48 logical layer visits per token. While weight reuse shrinks the stored checkpoint, the model still executes all 48 visits at runtime, maintaining separate key-value (KV) and short-convolution states for every pass.

Each pass through layers 6 through 17 begins from the output of the initial six-layer block, passing state across iterations via cross-pass messages. Every looped layer reads a message from the previous pass's outputs using a learned query—with separate messages for query, key, value, and MLP inputs—and applies a gating controller to balance retained input against incoming cross-pass information. During passes two and three, core layers attend over two residual blocks from the prior iteration, using four rank-64 gates to mix messages into the attention and MoE paths.

Inside each six-layer block, five layers apply sliding-window attention limited to a 1,024-token local window, while one layer performs full attention equipped with dynamic key and value convolutions. Attention operates with 24 query heads and 8 KV heads, incorporating learned relative position logits, a learned sink, and an output gate. The text decoder width stands at 2,304—half the 4,608 hidden size of Inkling 2 Small—and each SwiGLU expert uses an intermediate width of 1,152. At every layer visit, the router selects 8 of 128 available experts per token, normalizing scores and scaling weights by 8 (compared with 8 of 384 experts on Inkling 2 Small).

Reported Benchmarks and Validation Claims

According to the Hugging Face announcement, Thinking Machines Lab reports that Inkling 2 Mini achieves state-of-the-art scores in its size class on the Artificial Analysis Intelligence Index v5 (AAv5). Across comparisons against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B, and Nemotron 3.5 Super VL, the authors state that Inkling 2 Mini leads in 32 of 48 evaluated benchmarks covering coding, autoresearch, agentic search, knowledge-work tool use, vision interpretation, reasoning, and safety alignment. In coding evaluations specifically, the post claims the model leads on FrontierCode v2.0, ties Qwen3.8-27B on Terminal-Bench 4.0, and trails Qwen3.8-27B on AutomationBench-AA v1.0.6. The blog authors also describe Inkling 2 Mini as comparable to Qwen 3.8 27B in informal local tests while emphasizing that further independent validation is needed.

Developer Workflows and Inference Ecosystem

To support immediate adoption across diverse hardware, Inkling 2 Mini launches with integration across Hugging Face transformers (via AutoModelForMultimodalLM and AutoProcessor), Llama App (llama.cpp), Apple MLX for Apple Silicon, and SGLang for day-zero datacenter serving. Developers can tune compute allocation per request through a reasoning_effort argument in apply_chat_template, ranging from none (0.0) and medium (0.7) up to the default high (0.9) or max (0.99), or by passing any floating-point value between 0.0 and 0.99, which gets injected as a numeric system instruction. Native tool calling is also integrated into the chat template, allowing structured function invocation such as external API lookups. On the experimental TML-d24-varloop-research checkpoint, developers can adjust loop_num_passes in text_config to one or two passes for faster generation.

Caveats, Hardware Realities, and Open Questions

Despite promising architectural efficiency, several practical constraints warrant scrutiny before deployment. First, although looping halves distinct layer weights on disk, executing 48 logical visits with per-visit KV and short-convolution state means long-context workloads approaching the 1-million-token limit still carry substantial memory and latency demands. Second, the variable-loop research model has not been evaluated in depth, and reducing passes below three trades output quality for speed. Third, the announcement blog post does not specify the software license governing weight redistribution, exact VRAM thresholds for NVFP4 and MXFP8 local execution, or independent third-party benchmark audits—details teams will need to verify directly on the model card prior to production adoption.

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-21.

Follow your curiosity.

Pass the signal.