Inkling: Enterprise AI
Thinking Machines Launches Inkling, an Open-Weight AI for Enterprise Excelling in Agentic Tasks
Anonymous
Key Points
1
Inkling exposes two distinct context windows: 256K context via the Tinker API and 1M context via HuggingFace, enabling longer reasoning before context refresh.
2
Market coverage notes Inkling stands out on AI-agent tasks, signaling strength in agentic capabilities relative to Western open-model peers.
3
Inkling achieves notable agentic performance metrics, including an Elo of 1238 on GDPval-AA v2 and 24% on 𝜏³-Banking, surpassing certain contemporaries in these benchmarks.
4
Inkling is token-efficient, averaging about 25,000 output tokens per Intelligence Index task, lower than several open-weight peers in comparable tests.
5
Inkling uses about a third of the tokens compared with Nvidia Nemotron 3 Ultra to achieve the same code performance, highlighting potential efficiency advantages in certain workloads.
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has launched its first public AI model. Called Inkling, it packs 975 billion total parameters and is available as an open-weight model — meaning developers can download, modify, and deploy it freely, according to Crypto Briefing.
Inkling targets enterprises that want to customize AI for their own needs, rather than rely on closed, one-size-fits-all models from OpenAI, Anthropic, or Google. MarketScreener reports the model is also positioned as a direct competitor to open models built in China.
A 975-Billion-Parameter Model Built for Flexibility
Inkling uses a mixture-of-experts architecture. That means only about 41 billion of its 975 billion parameters are active at any one time, keeping it fast and efficient. It was trained on 45 trillion tokens across text, images, audio, and video, giving it broad multimodal ability, OfficeChai reports.
Developers can access Inkling in two ways. The Tinker API offers a 256,000-token context window. HuggingFace offers an even larger 1 million-token window — meaning the model can read and reason over far more information before needing a reset, according to Guru Focus.
Benchmark Results Put Inkling Ahead of Western Open Models
Inkling claims a strong lead in agentic tasks — work where AI takes a series of actions to complete a goal. It scored an Elo of 1,238 on GDPval-AA v2 and hit 24% on τ³-Banking, outpacing several comparable open-weight models on those tests, per OfficeChai.
Token efficiency is another standout. Inkling averages about 25,000 output tokens per Intelligence Index task. To match Nvidia's Nemotron 3 Ultra on code tasks, Inkling uses roughly one-third the tokens. Fewer tokens means faster, cheaper responses, MarketScreener UK notes.
Enterprise Customization Is the Core Selling Point
Thinking Machines argues that general-purpose AI leaves a gap for businesses with specialized needs. Inkling is designed to let organizations train the model on their own internal knowledge and workflows. That pitch sets it apart from locked, API-only services, according to Crypto Briefing.
The model also signals when it is uncertain — a reliability feature aimed at enterprise users who cannot afford confident wrong answers. It can also adjust its "thinking volume," trading depth of reasoning for speed depending on the task at hand.
Eighteen Months of Stealth Work Leads to Public Debut
Thinking Machines Lab spent roughly 18 months building in stealth before this release. Murati left OpenAI in late 2023 and founded the lab shortly after. Inkling is the first model the company has shared publicly, Crypto Briefing reports.
The open-weight release puts pressure on rivals who keep their model weights private. By letting developers inspect and modify Inkling directly, Thinking Machines is betting that openness — not secrecy — is the better path to enterprise adoption, MarketScreener notes.
Sentiment
Mostly PositiveNEGATIVENEUTRALPOSITIVE
Publishers
21
Articles
8
Reach
29
A
AnonymousAuthor