Overview
Inkling is Thinking Machines Lab's first foundation model, released on July 15, 2026. It is a sparse Mixture-of-Experts transformer with 975B total parameters and 41B active per token, trained on roughly 45 trillion tokens of text, images, audio, and video.
The model reasons natively over text, images, and audio (audio is ingested as dMel spectrograms, images as 40×40-pixel patches via an hMLP encoder), supports a 1M-token context, and exposes a controllable thinking-effort dial between 0.2 and 0.99 so users can trade token spend against answer quality.
Full weights are published on Hugging Face under Apache-2.0 (BF16 and NVFP4 checkpoints, the latter tuned for NVIDIA Blackwell). A smaller Inkling-Small preview (276B total · 12B active) uses the same recipe at lower cost. Thinking Machines positions Inkling explicitly as "not the strongest overall model available today" — it is meant as a customization base for the company's Tinker fine-tuning platform rather than a leaderboard-topping frontier model.
| Released | 2026-07-15 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 975B total · 41B active |
| Context | 1M |
| Architecture | Sparse Mixture-of-Experts (256 routed + 2 shared experts, 6 active per token) |
| Modalities | Text, Vision, Audio |
| Status | Generally available |
Benchmarks
Inkling benchmark comparison as published by Thinking Machines on the Inkling launch page (all Inkling evaluations at effort=0.99).
| Benchmark | Inkling | Nemotron 3 Ultra | Kimi K2.6 | GLM 5.2 | DeepSeek V4 Pro | Gemini 3.1 Pro | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|---|---|---|
| HLE (text only) | 29.7% | 26.6% | 35.9% | 40.1% | 35.9% | 44.7% | 53.3% | 47.2% |
| HLE (with tools) | 46% | 37.4% | 54% | 54.7% | 48.2% | 51.4% | 64.5% | 55% |
| AIME 2026 | 97.1% | 94.2% | 96.4% | 99.2% | 96.7% | 98.3% | 99.9% | 99.9% |
| GPQA Diamond | 87.2% | 86.7% | 91.1% | 89.5% | 88.8% | 94.1% | 92.6% | 94.1% |
| SWE-Bench Verified | 77.6% | 70.7% | 80.2% | 80% | 80.6% | 80.6% | 95% | 82.2% |
| Terminal Bench 2.1 | 63.8% | 56.4% | 71.3% | 82.7% | 64% | 73.8% | 84.6% | 89.5% |
| MCP Atlas | 74.1% | 44.7% | 68.1% | 77.8% | 73.2% | 78.2% | 83.3% | 81.8% |
| Global-MMLU-Lite | 88.7% | 85.6% | 88.4% | 89.2% | 89.3% | 92.7% | 93.3% | 91.8% |
| FORTRESS (Adversarial) | 78% | 77.6% | 65.6% | 71.3% | 36% | 65.2% | 96% | 82.4% |
| StrongREJECT | 98.6% | 98.7% | 99.8% | 98.5% | 98.6% | 98% | 98.7% | 98.5% |
This model's scores
- AIME 202697.1%
- GPQA Diamond87.2%
- SWE-Bench Verified77.6%
- HLE (with tools)46%
- MMMU Pro73.5%
- VoiceBench91.4%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Fully open weights under Apache-2.0, downloadable and commercially usable without a bespoke license
- Natively multimodal — reasons over text, images, and audio in a single pass
- 1M-token context window with a controllable 0.2–0.99 thinking-effort knob
- Sparse MoE (41B active out of 975B) — designed for efficient inference relative to dense frontier peers
- Blackwell-ready NVFP4 checkpoint published alongside the BF16 weights
Best for
- Reach for it as an open-weight base to fine-tune for a domain, product, or private data on the Tinker platform.
- Reach for it when you need a multimodal open-weight model with audio input, not just text and vision.
- Reach for it when a permissive Apache-2.0 license and self-hostable weights are hard requirements.
How to access
| Provider | Model ID |
|---|---|
| Together AI ↗ | — |
| Fireworks AI ↗ | — |
| Modal ↗ | — |
| Databricks ↗ | — |
| Baseten ↗ | — |
FAQ
What is Inkling?
Inkling is Thinking Machines Lab's first foundation model, released on July 15, 2026. It is a sparse Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per token, supporting text, image, and audio input with a 1M-token context window.
What license does Inkling use?
Inkling is released under the Apache-2.0 license. The full BF16 weights and an NVFP4 checkpoint tuned for NVIDIA Blackwell are published on Hugging Face at thinkingmachines/inkling and can be downloaded, fine-tuned, and commercialized without additional restrictions.
Where can I run Inkling?
Inkling is served on hosted APIs by Together AI, Fireworks AI, Modal, Databricks, and Baseten, or you can download the Apache-2.0 weights from Hugging Face and self-host. Thinking Machines also offers it as a fine-tuning base on its own Tinker platform.
Is Inkling a frontier model?
Thinking Machines explicitly states that Inkling is "not the strongest overall model available today, open or closed." Its published benchmark comparison places it below Claude Fable 5, GPT-5.6 Sol, and Gemini 3.1 Pro on most tasks. The model is designed as a customization base rather than a leaderboard-topping frontier model.