Thinking Machines Inkling Release Notes / What's New

Thinking Machines has released Inkling, a massive 1 trillion parameter open model capable of natively processing image, text, and audio inputs. This release is significant because it provides a large-scale, multimodal open model with a 1 million token context window and agentic reasoning capabilities.

Model Architecture and Capabilities

Inkling is a decoder-only multimodal Mixture-of-Experts (MoE) model. It utilizes a sparse architecture with 975 billion total parameters, of which 41 billion are active during any single inference pass, enabling faster computation.

Key Architectural Features

• Relative Attention: Instead of using Rotary Positional Embeddings (RoPE), Inkling uses relative attention where each attention layer learns position directly in the attention logits via a per-token, per-head relative feature projection.
• Hybrid Attention: The model employs a hybrid scheme alternating between global attention and sliding window attention in a 5:1 ratio, with the final layer utilizing global attention for rich feature representation.
• Short Convolution (SConv): A 1D convolution is applied over hidden states to handle local representations, reducing the computational load on the attention and MoE modules.
• MoE with Shared Experts: The router performs top-k selection over 256 experts, selecting 6 routed experts and 2 shared experts that remain always active.

Multimodal Processing

Unlike models that rely on separate encoders for each modality, Inkling uses relatively simple multimodal towers:

• Vision: A hierarchical MLP patchifier merges pixels progressively into embeddings. The architecture supports a temporal dimension for video processing.
• Audio: Audio is converted via a discretized mel spectrogram, where 100ms chunks are classified into mel spectrogram bins and then embedded.

Inference and Deployment

Inkling is available in two primary formats: a full BF16 checkpoint requiring 2 TB of VRAM, and a well-calibrated NVFP4 variant requiring 600 GB of VRAM.

Supported Frameworks

• Transformers: Day-0 support is available via the any-to-any pipeline in the transformers library (version 5.14.0+). Users can specify reasoning_effort levels ranging from none to max.
• High-Performance Engines: The model is supported in SGLang (which includes a custom implementation for speed), vLLM (optimized for production serving), and llama.cpp (for local deployment via GGUF quants).
• Local Deployment: Unsloth has provided 1-bit precision quantizations, which reduce VRAM consumption by 95% compared to the original model.

Performance and Benchmarks

Inkling demonstrates competitive performance across reasoning, agentic, and multimodal benchmarks. While it shows high proficiency in many areas, performance varies depending on the specific task and reasoning effort level.

Reasoning and Factuality

In text-only reasoning (HLE), Inkling scores 29.7%, trailing behind models like Claude Fable 5 (53.3%) and GPT 5.6 Sol (47.2%). However, it shows strong mathematical reasoning on AIME 2026 with a score of 97.1%.

Agentic and Multimodal Performance

• Coding: On SWEBench Verified, Inkling achieves 77.6%, performing competitively with models like Gemini 3.1 Pro (80.6%) and Claude Fable 5 (95.0%).
• Vision: On MMMU Pro (Standard 10), Inkling scores 73.3%.
• Audio: On VoiceBench, Inkling achieves a score of 91.4%.

Post-Training and Research Tools

Thinking Machines provides tools for developers looking to adapt the model through fine-tuning or reinforcement learning:

• Tinker: A managed tool for post-training open-weight models, including cookbooks for distillation and RL.
• OpenEnv: An agentic RL environment tool used to train the model using the ECHO algorithm, which allows a model to predict an environment without a separate verifier.
• Knowledge Distillation: Inkling can serve as a teacher model for smaller, on-device models using the GOLD algorithm to match token logits across different tokenizers.