Inkling Open-Weights Model Release

Thinking Machines has released Inkling, a multimodal Mixture-of-Experts (MoE) model designed specifically to serve as a flexible, open-weights foundation for customization and fine-tuning. Rather than competing for the top spot on general benchmarks, Inkling prioritizes a balance of multimodal capabilities, efficient "thinking" effort, and ease of adaptation for specialized real-world workflows.

Model Architecture and Specifications

Inkling is a Mixture-of-Experts transformer featuring 975 billion total parameters, with 41 billion active parameters per token. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens of text, images, audio, and video.

Technical Design

  • MoE Structure: The architecture follows a design similar to DeepSeek-V3, utilizing 256 routed experts and 2 shared experts, with 6 routed experts active per token.
  • Attention Mechanism: The model interleaves sliding-window and global layers at a 5:1 ratio with 8 KV heads. It uses relative positional embeddings (based on Shaw et al. and Huang et al.) instead of Rotary Positional Embedding (RoPE) to improve extrapolation for longer sequences.
  • Multimodality: Inkling uses an encoder-free architecture for audio and vision. Audio is processed as discrete dMel spectrograms, and images are encoded as 40x40 pixel patches using a four-layer hMLP. Both are processed jointly with text tokens.
  • Optimization: Training employed a hybrid strategy using Muon for large matrix weights and Adam for other parameters, with weight decay coupled to the square of the learning rate to maintain weight stability.

Controllable Thinking Effort

One of Inkling's primary differentiators is its support for controllable thinking effort, allowing developers to balance performance against latency and token cost.

By adjusting the effort setting (ranging from 0.2 to 0.99), users can control how many tokens the model spends on reasoning. Benchmarks on Terminal Bench 2.1, HLE, and IFBench indicate that Inkling can match the performance of other open-weights models, such as Nemotron 3 Ultra, while using significantly fewer tokens (approximately one-third in some cases).

Capabilities and Benchmarking

Inkling is positioned as a generalist model, trained across agentic tasks, reasoning, coding, and multimodal inputs to avoid narrow optimization.

Agentic Coding and Tool Use

Inkling is designed to operate within coding and agent harnesses. It has demonstrated the ability to:

  • Build functional web apps in a single shot.
  • Create multi-page, cohesively styled artifacts (e.g., a nine-page PDF journal).
  • Sustain long refinement loops, such as improving a multiplayer game through 40 iterations of feedback.

Multimodal Performance

Inkling reasons natively over text, images, and audio. It performs competitively on VoiceBench, MMAU, and AudioMC for audio, and shows strong results on charts, diagrams, and mathematical visual reasoning tasks. It can also leverage a Python tool during inference to perform image operations like zooming and cropping.

Epistemics and Safety

  • Calibration: The model was trained using RL against proper scoring rules to ensure it expresses appropriate confidence levels, particularly for forecasting tasks.
  • Instruction Following: Thinking Machines used a dual-grader system (a rubric grader for recall and a factuality grader for claim verification via agentic web search) to reduce hallucinations.
  • Censorship: The model was trained to answer directly on topics often subject to censorship, showing strong patterns of non-compliance with censorship constraints in evaluations by the Cognition team.
  • Safety: On the FORTRESS benchmark, Inkling demonstrated the strongest built-in safeguards among compared open-weights models, refusing harmful requests without over-refusing benign ones.

Inkling-Small Preview

Thinking Machines also previewed Inkling-Small, a 276B parameter MoE model with 12B active parameters. Despite its smaller size, Inkling-Small matches or exceeds the larger model on several benchmarks, including GPQA Diamond and IFBench, making it a candidate for low-latency workloads such as synthetic data generation or automated grading.

Ecosystem and Customization

Inkling is designed to be fine-tuned via the Tinker platform. To support this, Thinking Machines has released:

  • Inkling Playground: A developer interface for chatting and testing the model.
  • TML-Renderer: A tool for sampling and post-training with chat templates and multimodal inputs.
  • Deployment Partners: Integration with Together, Fireworks, Modal, Databricks, Baseten, SGLang, vLLM, TokenSpeed, llama.cpp, and Hugging Face transformers.

Community Insights and Analysis

Discussion among technical users highlights several key perspectives on the release:

  • Strategic Positioning: Users noted that Thinking Machines is competing on customizability rather than leaderboard dominance. One observer mentioned:

"Open base models that can be fine tuned on Tinker is a great business model... You (i.e. an enterprise) can own your own model & have it perform frontier-or-better at your task at potentially much lower cost."

  • Regional Competition: Some users viewed the release as a strong American entry into the open-weights space, which has recently been dominated by Chinese models like DeepSeek and GLM.

  • Cautions on Benchmarks: Some critics pointed out that the use of radar plots or specific benchmarking choices can obscure raw performance, urging developers to perform their own evaluations within their specific harnesses.

Sources

Related