Inkling
Mira Murati's Lab Launches a 975B Open Multimodal Model
Benchmarks & Evaluation Results
Reasoning
Matches or exceeds larger MoE models on multi-step reasoning tasks
Agentic Tasks
Strong performance on tool-use and autonomous coding benchmarks
Multimodal
Competitive with frontier models on vision-language evaluations
Inkling-Small
Matches Inkling on many benchmarks despite 3.5x fewer active params
Full Analysis of Inkling
Inkling represents Thinking Machines Lab's vision for open, customizable AI. Founded in February 2025, the company has been deliberately deliberate about its first model release — publishing novel research, building a customization platform, and previewing collaborative AI systems before shipping Inkling. The model itself is designed as a broad, balanced foundation model rather than a narrow benchmark optimizer. It was trained on an unusually diverse dataset covering text, images, audio, and v
ideo across 45 trillion tokens, with the goal of creating a flexible base that can be fine-tuned for specialized applications. The MoE architecture activates only 41 billion of its 975 billion parameters per token, making inference practical on clusters of 4-8 GPUs. The speculative MTP layers predict multiple future tokens simultaneously, improving generation speed without sacrificing quality. Thinking Machines has made Inkling fully available on Hugging Face under an open weights model, with bo
th BF16 and NVFP4 quantized formats. The company is also offering hosted inference on its Tinker platform at a 50% launch discount. Inkling-Small, the 12B-active-parameter variant, is particularly interesting: because of improvements in training data quality and recipe, it matches Inkling on many evaluations while being dramatically cheaper to run.
Strengths & Considerations
Strengths
- + Native any-to-any multimodal (text + image + audio) without separate adapters
- + Fully open weights on Hugging Face — download, fine-tune, self-host
- + Day-0 support in all major inference engines (transformers, SGLang, vLLM, llama.cpp)
- + Speculative MTP layers for faster inference throughput
- + Two size options for different cost targets (41B and 12B active)
- + NVFP4 quantization for efficient Blackwell GPU inference
- + Backed by Mira Murati's team with deep transformer expertise
Considerations
- − Trails top proprietary models on some frontier benchmarks
- − Inkling-Small still in preview; full release pending
- − 50% launch discount is time-limited
- − Smaller ecosystem and community compared to Llama or Qwen families
- − 975B total parameters still requires multi-GPU setup for self-hosting
Best Use Cases for Inkling
Where this model excels and the types of workloads it is best suited for.
Use Case 1
Multimodal applications needing native image, audio, and text understanding
Use Case 2
Self-hosted deployments for data-sensitive workloads
Use Case 3
Fine-tuning and customization for specialized domains
Use Case 4
Research requiring full model access and modification
Use Case 5
Synthetic data generation across multiple modalities
Use Case 6
Applications needing controllable thinking effort vs. cost tradeoffs
How to Use Inkling
Quick Start Guide
Download Inkling from Hugging Face at huggingface.co/thinkingmachines/Inkling. Install transformers v5.14.0+ (pip install -U transformers) and use the any-to-any pipeline for multimodal inference. For hosted access, visit Tinker at thinkingmachines.ai — currently at 50% launch discount. For self-hosting, the BF16 checkpoint requires Hopper-class GPUs; the NVFP4 checkpoint is optimized for Blackwell. SGLang and vLLM both support Inkling with day-0 integration.
Other Models from July 2026
Compare Inkling with other recent releases.