Thinking Machines' first model, Inkling, lands behind China's best

2 months ago

Mira Murati's lab raised what was reported as the largest seed round in history. Its first model, Inkling, just shipped, and the announcement itself says it is not the strongest model available today. That sentence is the story.

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Thinking Machines' first model, Inkling, lands behind China's best
  2. 0:14Audio and images go straight into the token stream.
  3. 0:26Their own table shows it trailing.
  4. 0:42Same score. A third of the tokens.
  5. 1:00It compressed its own chain of thought.
  6. 1:13Full weights on Hugging Face, day one.
  7. 1:27Their words: not the strongest.
Show transcript

Thinking Machines' first model, Inkling, lands behind China's best

Inkling: a 975B open-weights model with a dial for how hard it thinks
open weights 975B MoE thinkingmachines.ai

Mira Murati's lab raised what was reported as the largest seed round in history. Its first model, Inkling, just shipped, and the announcement itself says it is not the strongest model available today. That sentence is the story.

Audio and images go straight into the token stream.

Text
tokens, straight in
Audio
dMel spectrogram frames
Images
40×40 px patches → 4-layer hMLP
one token stream
no vision or audio encoder towers
975B total → 41B active
256 routed + 2 shared experts
6 active per token
975B total → 41B active 1M context 45T pretraining tokens
thinkingmachines.ai

It takes audio and images raw, straight into the token stream, no separate encoders. Nine hundred and seventy five billion parameters, forty one billion active per token. And the full weights are downloadable now.

Their own table shows it trailing.

the lab reports
Inkling Nemotron 3 Ultra Kimi K2.6 GLM 5.2
HLE (text only) 29.7 26.6 35.9 40.1
Terminal Bench 2.1 63.8 56.4 71.3 82.7
SWEBench Verified 77.6 70.7 80.2 80.0
FORTRESS Adversarial 78.0 77.6 65.6 71.3
Independent index: 41 — top Western open weights (GLM 5.2: 51) · artificialanalysis.ai
thinkingmachines.ai · artificialanalysis.ai

The numbers back the modesty. The lab's own table has it trailing GLM 5.2 on most raw scores. An independent index rates it the strongest Western open weights model, still behind the Chinese leader, despite its size. Then why ship it?

Same score. A third of the tokens.

Terminal Bench 2.1 · score % 20 40 60 80 0 25k 50k 75k 100k mean generated tokens Inkling · effort 0.2 → 0.99
Kimi K2.5 · 18.7k Nemotron 3 Ultra · 32.8k Kimi K2.6 · 42.3k GLM 5.2 · 82.1k 11.9k vs 32.8k tokens
the lab reports; competitors at their default operating point
thinkingmachines.ai

The answer is a dial for how hard it thinks. On one benchmark, the lab reports, it matches Nemotron 3 Ultra at roughly a third of the tokens. Not the strongest model. The cheapest path to yours. The weights are free, because the fine tuning platform, Tinker, is the product.

It compressed its own chain of thought.

30M+ RL rollouts
Early in RL — verbose, grammatical
“We need to understand the operator.”
“The internal coordinate is periodic.”
“So the internal metric is e^{2A(x)} dx².”
Late in RL — compressed, telegraphic
“We need determine eigenvalue problem for spin-2 fluctuations…”
“Wait internal metric is e^{2A} dx²?”
“Yes.”
thinkingmachines.ai

Across more than thirty million reinforcement learning rollouts, the model compressed its own chain of thought. Early traces read like prose. Late ones read like a telegram. Nobody asked it to. Saving tokens alone drove it.

Full weights on Hugging Face, day one.

Hugging Face model card: thinkingmachines/Inkling
Apache-2.0 + acceptable-use policy huggingface.co · thinkingmachines.ai

The weights are on Hugging Face, original and quantized, Apache licensed with an acceptable use policy attached. Day one support spans the major hosts and inference stacks. A smaller sibling is previewed, not yet released.

Their words: not the strongest.

Their bet: cost per result.
thinkingmachines.ai

Not the strongest model. Their words. The bet is that a tunable, open base is worth more than a leaderboard crown. Watch what gets built on it.