Thinking Machines open sources first multimodal language model, Inkling, focused on low cost and 'resistance to censorship'

Summary

Thinking Machines released Inkling, an open-source multimodal language model designed for cost-efficiency and censorship resistance. It performs well on benchmarks, though some Chinese models and closed-source alternatives lead in specific areas. Inkling's unique 'controllable thinking effort' allows developers to balance cost and performance.

Key Takeaways

1. Thinking Machines released Inkling, its first major language model under an Apache 2.0 open source license, designed for enterprises seeking customizable, controllable, and on-premises AI solutions.
2. Inkling is a natively multimodal, open-weights Mixture-of-Experts (MoE) system capable of reasoning across text, images, and audio, with 975 billion total parameters.
3. Inkling was designed 'to answer directly on topics that may be subject to censorship,' offering enterprises concerned about factual outputs a more trustworthy option.

Statement Breakdown

• Claimed Facts: 60% of statements the article presents as facts
• Opinions: 30% of statements classified as editorial or subjective
• Claims: 10% of statements surfaced for additional reader evaluation

Credibility & Bias Reasoning

Credibility assessment: The article presents a balanced view by comparing Inkling's performance against various open-source and closed-source models, citing specific benchmarks. It acknowledges limitations and potential vulnerabilities, offering practical advice for developers. The information is largely factual and well-supported by benchmark data.

Bias assessment: Tech Enthusiast. The article's tone is enthusiastic about AI advancements and the potential of open-source models. It highlights the innovative aspects of Inkling and its competitive positioning within the AI landscape, framing it as a significant development.

Note: This article provides a detailed technical overview of a new AI model. While it cites benchmarks, consider the inherent excitement surrounding new AI releases and seek independent verification for performance claims.

Credibility flag: Informative, but watch for hype

Claimed Facts (10)

• This statement sets the context for the article by describing a market need that the new model addresses.
• This provides specific details about the model's release, licensing, and benchmark performance, presenting factual data.
• This statement provides concrete technical specifications of the Inkling model.
• This states where the model's weights and API can be accessed, providing factual information.
• This announces a secondary product offering with specific parameter count, presenting factual information.
• This presents a comparative performance analysis based on benchmark data.
• This provides specific benchmark comparisons between Inkling and another model.
• This offers a direct comparison of benchmark scores between Inkling and Kimi K2.6.
• This presents a comparative performance statement against a specific competitor.
• This provides specific benchmark data for a multimodal capability, comparing it to other models.

Opinions (10)

• This statement frames the release as a 'strong new contender,' which is an evaluative opinion.
• The term 'more trustworthy option' is a subjective assessment of the model's benefit.
• Calling it a 'significant departure' is an interpretation of its architectural approach.
• Describing the landscape as 'fiercely competitive' and Inkling as 'formidable' are subjective assessments.
• Characterizing the model as a 'broad, balanced generalist' is an interpretation of its design philosophy.
• The phrase 'stiff challenge' and the implication of 'ultimately outperforming' are evaluative statements.
• The phrase 'maintain the edge' is a subjective assessment of competitive advantage.
• Claims of being the 'most capable' and having a 'highly defensible position' are strong subjective assertions.
• Calling a feature 'standout' is a subjective judgment of its importance.
• The characterization of other models as having 'overly restrictive safety guardrails' or echoing 'state-aligned ideological talking points' is an opinionated framing.

Claims (5)

• While Mira Murati was formerly at OpenAI, she is not the CTO. This is a factual inaccuracy that casts doubt on the article's attention to detail.
• The claim of 'strong patterns of censorship non-compliance' and 'effectively resisting ideological capture' is based on a specific evaluation by a startup (Cognition) and could be subject to interpretation or bias in the evaluation itself.
• This statement, while presented as a finding, describes a potential vulnerability that is not fully quantified or explained, making it a somewhat vague and potentially downplayed concern.
• While practical advice, this implicitly suggests the model's built-in safety is insufficient, which could be seen as a subtle way to manage expectations or highlight a known weakness without fully detailing its severity.
• The claim of 'reaching the same score with a fraction of the tokens' is a strong performance assertion that may require more detailed substantiation beyond a general statement.

Key Sources

• Thinking Machines — AI Startup
• Mira Murati — Former OpenAI CTO
• Nvidia — Technology Company
• Google — Technology Company
• Hugging Face — AI Community Platform
• VentureBeat — Technology News Outlet
• GLM — AI Model
• DeepSeek — AI Lab
• Kimi — AI Model
• OpenAI — AI Research Lab
• Anthropic — AI Safety Company
• Cognition — AI Startup