How do I actually run Inkling?

Inkling weights are on HuggingFace as BF16 (needs ~2 TB of VRAM) or a quantized NVFP4 checkpoint (~600 GB). Thinking Machines lists Together, Fireworks, Modal, Databricks, and Baseten as launch hosting partners, and its own Tinker platform runs both a free playground and paid fine-tuning.

Is Inkling the strongest open model right now?

No — Thinking Machines states plainly that Inkling is 'not the strongest model available today, closed or open.' It ships reported scores of 77.6% on SWE-bench Verified, 87.2% on GPQA Diamond, and 97.1% on AIME 2026, but the pitch is customizability and multimodal reasoning, not topping every leaderboard.

What's actually new architecturally in Inkling?

Inkling is a 66-layer decoder-only transformer that routes each token to 6 of 256 experts plus 2 shared experts. Attention interleaves sliding-window and global layers at a 5:1 ratio, uses relative positional embeddings instead of RoPE, and adds short convolutions in the attention and residual paths. Training combined Muon (large matrices) with Adam.

Why did Thinking Machines pick 'open weights' for its first model?

Thinking Machines is betting that enterprises want a base model they can fine-tune into a private specialist rather than rent from a frontier lab. Inkling is Apache 2.0, downloadable, and paired with Tinker — a hosted fine-tuning stack that turns the open weights into recurring platform revenue.

How does Inkling handle images and audio?

Inkling accepts UTF-8 text, pixel-based images (roughly 40–4096 px per side), and 16 kHz WAV audio up to about 20 minutes. Audio is discretized to dMel spectrograms and interleaved with text and image patches, letting a single decoder reason across all three modalities. Output stays text-only.

Thinking Machines Lab · 2026-07-15 · major

Inkling — Thinking Machines' first open-weights 975B/41B multimodal MoE

Thinking Machines' first public foundation model: a 975B / 41B-active MoE released under Apache 2.0. Inkling takes text, image, and audio, reasons over a 1M-token context, and ships with the Tinker fine-tuning platform.

Thinking Machines Lab 'Introducing Inkling' announcement cover.

Mira Murati's Thinking Machines Lab ships its first foundation model, and puts the full weights on HuggingFace under Apache 2.0.

Key specs

Active params	41B / 975B total
Context window	1M tokens
SWE-bench	77.6%
GPQA	87.2%
Aime 2026	97.1%

Quick facts

Maker	Thinking Machines Lab (Mira Murati)
Size	975B total, 41B active (Mixture-of-Experts)
License	Apache 2.0 (open weights)
Modalities in	Text, image, audio
Modalities out	Text only
Context window	1M tokens (64K / 256K on Tinker)
Training	45T tokens on NVIDIA GB300 systems, Muon + Adam optimizers
Availability	HuggingFace weights, Tinker playground, Together / Fireworks / Modal / Databricks / Baseten APIs

What is it?

Inkling is a 975-billion-parameter Mixture-of-Experts model that activates 41 billion parameters per token. Text, images, and audio go in; text comes out. It runs on a 1-million-token context and is Thinking Machines Lab's first public foundation model — the piece of the company plan that follows the Tinker fine-tuning platform and TML-Interaction-Small voice model.

How does it work?

Each Inkling forward pass routes tokens through 6 of 256 experts (plus 2 shared experts) across 66 decoder layers, with sliding-window and global attention interleaved at a 5:1 ratio and short convolutions added to the residual path. Training used 45 trillion multimodal tokens on NVIDIA GB300 systems, combining the Muon optimizer for the large matrices with Adam elsewhere, plus over 30 million reinforcement-learning rollouts to shape reasoning behaviour.

Why does it matter?

Inkling is the first open-weights frontier-scale release from Thinking Machines, and it lands as a customization play rather than a leaderboard bid — the company says outright that Inkling is not the strongest model available. Enterprises get a base they can legally fine-tune, and Thinking Machines gets a business (Tinker) around adapting it. That is a different bet from OpenAI, Anthropic, and Google Gemini, all of which keep frontier weights private.

Who is it for?

ML researchers, enterprise ML teams, and open-weights users who want a multimodal MoE base they can fine-tune.

Frequently asked questions

How do I actually run Inkling?: Inkling weights are on HuggingFace as BF16 (needs ~2 TB of VRAM) or a quantized NVFP4 checkpoint (~600 GB). Thinking Machines lists Together, Fireworks, Modal, Databricks, and Baseten as launch hosting partners, and its own Tinker platform runs both a free playground and paid fine-tuning.
Is Inkling the strongest open model right now?: No — Thinking Machines states plainly that Inkling is 'not the strongest model available today, closed or open.' It ships reported scores of 77.6% on SWE-bench Verified, 87.2% on GPQA Diamond, and 97.1% on AIME 2026, but the pitch is customizability and multimodal reasoning, not topping every leaderboard.
What's actually new architecturally in Inkling?: Inkling is a 66-layer decoder-only transformer that routes each token to 6 of 256 experts plus 2 shared experts. Attention interleaves sliding-window and global layers at a 5:1 ratio, uses relative positional embeddings instead of RoPE, and adds short convolutions in the attention and residual paths. Training combined Muon (large matrices) with Adam.
Why did Thinking Machines pick 'open weights' for its first model?: Thinking Machines is betting that enterprises want a base model they can fine-tune into a private specialist rather than rent from a frontier lab. Inkling is Apache 2.0, downloadable, and paired with Tinker — a hosted fine-tuning stack that turns the open weights into recurring platform revenue.
How does Inkling handle images and audio?: Inkling accepts UTF-8 text, pixel-based images (roughly 40–4096 px per side), and 16 kHz WAV audio up to about 20 minutes. Audio is discretized to dMel spectrograms and interleaved with text and image patches, letting a single decoder reason across all three modalities. Output stays text-only.

Try it

https://tinker.thinkingmachines.ai