Thinking Machines Lab drops Inkling & Meta’s Muse Spark 1.1

2026-07-17 · Source: IBM Technology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

Thinking Machines Lab launched Inkling, a 975B total parameter, 41B active Mixture of Experts (MoE) model. This open-source, multimodal model emphasizes customization and speed, supported by its Tinker fine-tuning platform. Concurrently, Meta released Muse Spark 1.1 from its Super Intelligence Lab, a benchmark-leading model with a million-token context window, strong coding harnesses, and cost-efficiency for agent workloads, targeting enterprise use cases. OpenAI's GPT 5.6 Soul achieved an 8% score on the challenging ARC AGI 3 benchmark, indicating progress in general intelligence. Anthropic's JSpace paper explored internal model reasoning, suggesting "subconscious" processing via attribution graphs, with implications for AI safety and control.

Key takeaway

For AI/ML Engineers evaluating new model deployments, recognize that raw benchmark scores are becoming less critical than a model's architectural design and ecosystem. Prioritize models offering robust fine-tuning capabilities, native multimodal integration, or advanced interpretability tools like JSpace. This shift allows for greater customization, cost-efficiency, and enhanced safety, aligning models more closely with specific enterprise requirements beyond general intelligence.

Key insights

The AI landscape is shifting from pure benchmark leadership to models emphasizing customization, efficiency, and interpretability.

Principles

• Open-source, customizable models challenge closed-source leaders.
• True multimodal architectures integrate data at the token level.
• Internal model states can be analyzed for safety and control.

Method

Thinking Machines' Tinker API enables automated, closed-loop fine-tuning. Anthropic's JSpace technique uses Jacobian-based analysis to light up internal model activations.

In practice

• Utilize Inkling for custom multimodal applications.
• Deploy Muse Spark 1.1 for economical agent workloads.
• Explore JSpace-like tools for agent safety and control.

Topics

• Mixture-of-Experts
• Multimodal AI
• Model Fine-tuning
• AI Benchmarking
• Model Interpretability
• Agentic AI
• AI Safety

Best for: CTO, VP of Engineering/Data, AI Architect, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Complete Guide to Thinking Machines Inkling — Inkling is an open-weights multimodal MoE model prioritizing customization and domain-specific fine-tuning over benchmark supremacy.
• What We are Missing in Multimodal LLM Evaluation? — Current MLLM evaluation benchmarks fail to assess cross-modal information integration, hindering true progress measurement.
• LWiAI Podcast #234 - Opus 4.6, GPT-5.3-Codex, Seedance 2.0, GLM-5 — The AI sector is rapidly advancing with new models, generative media tools, and significant venture capital investments.
• Together AI expands fine-tuning service with tool calling, reasoning, and vision support — Together AI's fine-tuning service now natively supports complex AI workflows like tool calling, reasoning, and VLM training.
• The Capability Frontier: Benchmarks Miss 82% of Model Performance — LLM benchmarks significantly understate collective capabilities due to single-model, single-run evaluations, missing 82% of potential performance.
• Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models — Multimodal LLMs exhibit significant order sensitivity, with current prompt-level mitigations proving insufficient for general robustness.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by IBM Technology.