thinkingmachines/Inkling-2-Mini · Hugging Face
Blog

thinkingmachines/Inkling-2-Mini · Hugging Face

2026.11.21
·Hugging Face·by Mineru
#AI#DL#LLM#ML#RAG

Key Points

  • 1Inkling 2 Mini is a 26.4B parameter multimodal looped transformer that utilizes a sparse Mixture-of-Experts architecture to deliver efficient performance for both text and image processing.
  • 2Designed for local deployment and agentic tasks, the model features a hybrid attention mechanism and runs middle layers three times to achieve increased computational depth with only 5.4B effective active parameters.
  • 3While the model demonstrates strong capabilities in coding and tool-use benchmarks, it exhibits known limitations such as potential factual inaccuracies, repetitive reasoning loops, and lower performance in non-English languages.

Inkling 2 Mini, developed by Thinking Machines Lab, is a general-purpose multimodal transformer model designed for deployment on local hardware. Released on November 20, 2026, under the Apache 2.0 license, it supports both text and image inputs to generate text, catering to applications such as coding assistants, agentic tool-use systems, and retrieval-augmented generation.

Core Architecture and Methodology

The model utilizes a 24-layer decoder-only transformer architecture enhanced with sparse Mixture-of-Experts (MoE) feed-forward layers. A distinguishing technical feature is its "looped" design: the middle 12 layers are executed three times with shared weights, effectively creating 48 layer applications per token from a 24-layer weight set.

In terms of parameterization, the model contains 26.4B total parameters, with 3.4B active parameters per single pass through the layers. Due to the three-pass looping mechanism, the effective active parameters per token reach approximately 5.4B. Within each layer, every token is routed to 8 out of 128 available experts. The model employs a hybrid attention mechanism combining local and global layers, projecting multimodal inputs into a shared hidden space.

The model is distributed as a single 20GB checkpoint utilizing mixed NVFP4 and MXFP8 numerical precisions. It supports a context window of up to 1M tokens.

Performance and Capabilities

Inkling 2 Mini demonstrates strong performance in agentic and coding-related benchmarks. It excels in tasks involving terminal interaction, SWEBench Verified (79.2%), and tool use/search-based tasks like BrowseComp (79.0%) and DeepSearchQA (85.3%). Despite its smaller parameter footprint, it remains competitive with larger models on technical reasoning and coding benchmarks, although it exhibits performance degradation in factual recall and knowledge-heavy domains compared to its larger counterpart, Inkling 2 Small.

Safety and Monitorability

A specific research focus for this model is the impact of weight-looping on monitorability. Because the model performs additional computation without explicitly generating chain-of-thought tokens, the researchers investigated whether this opaque reasoning process could be overseen. Their findings indicate that while the model has limited controllability over its internal reasoning, it failed to demonstrate the ability to perform covert side tasks in agentic environments.

Safety evaluations involved adversarial testing via the FORTRESS and StrongREJECT frameworks. The model includes safety mitigations derived from on-policy distillation. The developers note that while the model exhibits lower rates of misbehavior and destructive action tendencies compared to larger baselines, it still possesses limitations regarding factual accuracy, repetitive reasoning loops, and potential for unconfirmed destructive file operations.

Deployment and Usage

Designed for accessibility, Inkling 2 Mini is compatible with local AI compute environments, requiring a minimum of 32GB of memory. Supported frameworks include Hugging Face, SGLang, vLLM, and llama.cpp. The model is also available for fine-tuning via the Tinker platform and for inference via specialized renderers. The developers emphasize that due to its susceptibility to hallucination and agentic unreliability, the model should be operated within sandboxed environments with human-in-the-loop oversight for high-stakes or safety-critical applications.