Thinking Machines Launches Open AI Model, Inkling, to Challenge Norms
WHO:
Wired - Thinking Machines Lab is the organization launching the AI model.
TechCrunch - The company is identified as Thinking Machines.
WHAT:
VentureBeat - Thinking Machines announced a preview of Inkling-Small, a lighter 276-billion-parameter alternative optimized for low latency and cost.
VentureBeat - Inkling achieved a score of 1257 on Design Arena’s Agentic Web Dev leaderboard, measuring human scores of frontend web design.
VentureBeat - Inkling scored 54.3% on SWEBench Pro (Public), compared to GLM 5.2's 62.1%.
VentureBeat - Inkling scored 63.8% on Terminal Bench 2.1, while GLM 5.2 scored 82.7%.
VentureBeat - Inkling scored 30.0% on HLE (text only), compared to GLM 5.2's 40.1%.
VentureBeat - Inkling scored 43.9% on SimpleQA Verified, while DeepSeek V4 Pro scored 57.0%.
VentureBeat - Inkling scored 87.9% on GPQA Diamond, while Kimi K2.6 scored 91.1%.
VentureBeat - Inkling scored 77.1% on BrowseComp, compared to Kimi K2.6's 83.2%.
VentureBeat - Inkling scored 74.1% on MCP Atlas, outperforming Nemotron's 44.7%.
VentureBeat - Inkling scored 77.2% on MMAU in audio processing, while Gemini 3.1 Pro scored 82.5%.
VentureBeat - Inkling achieved a 97.1% score on AIME 2026 for mathematical problem-solving, surpassing DeepSeek's 96.7%.
VentureBeat - Inkling scored 79.8% on IFBench for general chat instruction following, outperforming Kimi K2.6's 76.0%.
VentureBeat - Inkling scored 91.4% on VoiceBench, compared to Gemini 3.1 Pro's 94.4%.
VentureBeat - Inkling achieved 77.6% on SWE-bench Verified, outperforming Nvidia Nemotron 3's 71.9%.
TechCrunch - Inkling is the first public proof point for Thinking Machines after a year and a half spent building AI infrastructure largely out of public view.
VentureBeat - Inkling is a natively multimodal, open-weights Mixture-of-Experts (MoE) system capable of reasoning across text, images, and audio.
VentureBeat - Inkling is positioned as a customization base rather than the strongest model available, emphasizing its controllable thinking effort as a practical differentiator.
VentureBeat - Inkling's weights are already available on Hugging Face and the company’s own model training application programming interface (API), Tinker.
VentureBeat - Inkling's architecture features a unique encoder-free early fusion approach for multimodal input.
VentureBeat - Inkling is released under the Apache 2.0 open source license, allowing developers to download, modify, integrate, and commercialize the model weights royalty-free.
VentureBeat - Inkling achieved a 78.0% refusal rate on adversarial queries while maintaining a 95.9% compliance rate on benign queries.
VentureBeat - Inkling supports a context window of 1 million tokens and uses relative positional embeddings instead of Rotary Positional Embedding (RoPE).
VentureBeat - Inkling's architecture features 975 billion total parameters, with only 41 billion active during token generation.
VentureBeat - Inkling is readily deployable across major open-source inference libraries, including SGLang, vLLM, TokenSpeed, and llama.cpp.
VentureBeat - Inkling was designed to be a broad, balanced generalist model rather than optimized for specific tasks.
VentureBeat - Inkling achieved a score of 1257 on Design Arena’s Agentic Web Dev leaderboard, measuring human scores of frontend web design.
VentureBeat - Inkling scored 54.3% on SWEBench Pro (Public), compared to GLM 5.2's 62.1%.
VentureBeat - Inkling scored 63.8% on Terminal Bench 2.1, while GLM 5.2 scored 82.7%.
VentureBeat - Inkling scored 30.0% on HLE (text only), compared to GLM 5.2's 40.1%.
VentureBeat - Inkling scored 43.9% on SimpleQA Verified, while DeepSeek V4 Pro scored 57.0%.
VentureBeat - Inkling scored 87.9% on GPQA Diamond, while Kimi K2.6 scored 91.1%.
VentureBeat - Inkling scored 77.1% on BrowseComp, compared to Kimi K2.6's 83.2%.
VentureBeat - Inkling scored 74.1% on MCP Atlas, outperforming Nemotron's 44.7%.
VentureBeat - Inkling scored 77.2% on MMAU in audio processing, while Gemini 3.1 Pro scored 82.5%.
VentureBeat - Inkling's "controllable thinking effort" allows developers to adjust the reasoning budget from 0.2 to 0.99.
VentureBeat - Inkling's training involved over 30 million rollouts during large-scale reinforcement learning.
VentureBeat - Inkling demonstrated "strong patterns of censorship non-compliance" in the Propaganda and Censorship Eval.
VentureBeat - Inkling achieved a 98.6% score on the StrongREJECT benchmark, aligning with strict frontier safety standards.
VentureBeat - Inkling's architecture features a unique encoder-free early fusion approach for multimodal input.
VentureBeat - Inkling is readily deployable across major open-source inference libraries, including SGLang, vLLM, TokenSpeed, and llama.cpp.
VentureBeat - Inkling was designed to answer directly on politically sensitive or heavily censored topics, demonstrating "strong patterns of censorship non-compliance."
VentureBeat - Inkling achieved a 78.0% refusal rate on adversarial queries while maintaining a 95.9% compliance rate on benign queries.
VentureBeat - Inkling scored 98.6% on the StrongREJECT benchmark, aligning with strict frontier safety standards.
VentureBeat - Inkling's architecture features 975 billion total parameters, with only 41 billion active during token generation.
VentureBeat - Inkling supports a context window of 1 million tokens and uses relative positional embeddings instead of Rotary Positional Embedding (RoPE).
VentureBeat - Inkling's "controllable thinking effort" allows developers to adjust the model's reasoning budget from 0.2 to 0.99.
VentureBeat - Inkling is released under the Apache 2.0 open source license, allowing developers to download, modify, integrate, and commercialize the model weights royalty-free.
VentureBeat - Inkling achieved 97.1% on AIME 2026 for mathematical problem-solving, surpassing DeepSeek's 96.7%.
VentureBeat - Inkling scored 79.8% on IFBench for general chat instruction following, outperforming Kimi K2.6's 76.0%.
VentureBeat - Inkling scored 91.4% on VoiceBench, compared to Gemini 3.1 Pro's 94.4%.
VentureBeat - Inkling achieved 77.6% on SWE-bench Verified, outperforming Nvidia Nemotron 3's 71.9%.
MarkTechPost - Inkling was released on July 15, 2026, as Thinking Machines' first model trained from scratch.
VentureBeat - Inkling is a 975-billion-parameter Mixture-of-Experts transformer with 41 billion active parameters, a 1 million-token context window, and native text, image, and audio input.
VentureBeat - Inkling is positioned as a customization base rather than the strongest model available, emphasizing its controllable thinking effort as a practical differentiator.
VentureBeat - Thinking Machines announced a preview of Inkling-Small, a lighter 276-billion-parameter alternative optimized for low latency and cost.
VentureBeat - Inkling was designed to answer directly on politically sensitive or heavily censored topics, demonstrating "strong patterns of censorship non-compliance."
VentureBeat - Inkling achieved a 97.1% score on AIME 2026 for mathematical problem-solving, surpassing DeepSeek's 96.7%.
VentureBeat - Inkling scored 79.8% on IFBench for general chat instruction following, outperforming Kimi K2.6's 76.0%.
MarkTechPost - Inkling is positioned as a customization base rather than the strongest model available, emphasizing its controllable thinking effort as a practical differentiator.
VentureBeat - Inkling is released under the Apache 2.0 open source license, allowing developers to download, modify, integrate, and commercialize the model weights royalty-free.
VentureBeat - Inkling achieved 77.6% on SWE-bench Verified, outperforming Nvidia Nemotron 3's 71.9%.
VentureBeat - Inkling scored 91.4% on VoiceBench, compared to Gemini 3.1 Pro's 94.4%.
VentureBeat - Inkling's "controllable thinking effort" allows developers to adjust the model's reasoning budget from 0.2 to 0.99.
VentureBeat - Inkling scored 98.6% on the StrongREJECT benchmark, aligning with strict frontier safety standards.
VentureBeat - Inkling achieved a 78.0% refusal rate on adversarial queries while maintaining a 95.9% compliance rate on benign queries.
VentureBeat - Inkling supports a context window of 1 million tokens and uses relative positional embeddings instead of Rotary Positional Embedding (RoPE).
VentureBeat - Inkling's architecture features 975 billion total parameters, with only 41 billion active during token generation.
Techmeme - Inkling is an open-weight MoE model with 975 billion total parameters and 41 billion active parameters, trained to be broad rather than optimized for one area.
Wired - The model, named Inkling, has 975 billion parameters.
Wired - Inkling is an open-source model designed to understand video and audio.
TechCrunch - Inkling is the first public proof point for Thinking Machines after a year and a half of development.
WHERE:
Wired - The launch is associated with the AI industry, competing against companies like Anthropic and OpenAI.
WHEN:
TechCrunch - The company spent a year and a half building AI infrastructure before launching Inkling.
WHY:
Techmeme - Thinking Machines Lab's mission is to build AI that extends human will and judgment.
Wired - The launch of Inkling aims to help Thinking Machines establish itself among competitors in the AI market.
HOW:
VentureBeat - Inkling was trained using a novel "controllable thinking effort" mechanism, allowing for a balance between cost and performance.
VentureBeat - Inkling was trained using a novel "controllable thinking effort" mechanism, allowing for a balance between cost and performance.
TechCrunch - The model was built largely out of public view, indicating a private development process.
(UPDATED July 16, 2026 18:19 UTC)
Thinking Machines has unveiled a preview of Inkling-Small, a lighter version with 276 billion parameters designed for low latency and cost efficiency. In various performance benchmarks, Inkling achieved impressive scores, including 97.1% on AIME 2026 for mathematical problem-solving, which surpassed DeepSeek's 96.7%. However, it faced tougher competition in other areas, scoring 54.3% on SWEBench Pro and 63.8% on Terminal Bench 2.1, compared to GLM 5.2's higher scores of 62.1% and 82.7%, respectively.
--
(UPDATED July 16, 2026 16:20 UTC)
Inkling has made its debut as the first public proof point for Thinking Machines after a year and a half of development in AI infrastructure. This innovative system is a natively multimodal, open-weights Mixture-of-Experts model that can reason across text, images, and audio, and it emphasizes customization over being the strongest model available. With a robust architecture featuring 975 billion total parameters and a context window of 1 million tokens, Inkling is now available under the Apache 2.0 open source license, allowing developers to freely download and modify its model weights.
--
(UPDATED July 16, 2026 15:27 UTC)
Inkling has emerged as a versatile generalist model, achieving a score of 1257 on Design Arena’s Agentic Web Dev leaderboard for frontend web design. In various benchmarks, Inkling scored 54.3% on SWEBench Pro, 63.8% on Terminal Bench 2.1, and 30.0% on HLE, while also demonstrating strong performance with 87.9% on GPQA Diamond and 74.1% on MCP Atlas. Notably, Inkling's architecture includes a unique encoder-free early fusion approach for multimodal input and is designed for easy deployment across major open-source inference libraries.
--
(UPDATED July 16, 2026 12:21 UTC)
Inkling has made headlines for its impressive performance on various benchmarks, achieving a remarkable 97.1% on AIME 2026 for mathematical problem-solving, which surpasses DeepSeek's score of 96.7%. The model also demonstrated a strong refusal rate of 78.0% on adversarial queries while maintaining a compliance rate of 95.9% on benign queries, showcasing its ability to navigate politically sensitive topics. Released under the Apache 2.0 open source license, Inkling features 975 billion parameters and allows developers to adjust its reasoning budget, making it a versatile tool for various applications.
--
(UPDATED July 16, 2026 01:21 UTC)
Thinking Machines has just released Inkling, its first model trained from scratch, which boasts an impressive 975-billion-parameter Mixture-of-Experts transformer. With 41 billion active parameters and a one million token context window, Inkling supports native text, image, and audio input, making it a versatile tool for various applications. This model emphasizes a "controllable thinking effort" mechanism, striking a balance between cost and performance while serving as a customization base rather than simply the strongest model on the market.
--
(UPDATED July 16, 2026 00:15 UTC)
Thinking Machines has unveiled a preview of Inkling-Small, a lighter alternative with 276 billion parameters that is optimized for low latency and cost. This innovative model is designed to tackle politically sensitive topics while demonstrating strong patterns of censorship non-compliance, and it achieved an impressive 97.1% score on AIME 2026 for mathematical problem-solving, surpassing DeepSeek's 96.7%. Additionally, Inkling scored 79.8% on IFBench for general chat instruction following, outperforming Kimi K2.6's 76.0%, positioning it as a customizable base rather than the strongest model on the market.
--
(UPDATED July 15, 2026 23:17 UTC)
Inkling has been released under the Apache 2.0 open source license, enabling developers to download, modify, and commercialize the model weights without any royalties. The model has shown impressive performance, achieving 77.6% on the SWE-bench Verified, which surpasses Nvidia Nemotron 3's score of 71.9%, and it scored 91.4% on VoiceBench, just shy of Gemini 3.1 Pro's 94.4%. With a robust architecture featuring 975 billion total parameters and a unique "controllable thinking effort" mechanism, Inkling allows developers to adjust the model's reasoning budget while maintaining a high compliance rate on benign queries.
--
(UPDATED July 15, 2026 20:26 UTC)
Inkling is an innovative open-weight mixture of experts model boasting a staggering 975 billion total parameters, with 41 billion of those being active at any given time. This model has been designed with a broad focus rather than being optimized for a specific area, showcasing the versatility of modern AI. The Thinking Machines Lab, which developed Inkling, aims to create artificial intelligence that enhances human will and judgment.
--
(UPDATED July 15, 2026 19:28 UTC)
Thinking Machines Lab has officially launched its new AI model named Inkling, which boasts an impressive 975 billion parameters and is designed to understand both video and audio. This launch marks the first public proof point for Thinking Machines after a year and a half of private development, positioning the company to compete in the AI industry against major players like Anthropic and OpenAI. Inkling is also an open-source model, reflecting the organization's commitment to transparency and innovation in the rapidly evolving tech landscape.