🗞️ Mira Murati's Thinking Machines Lab drops massive open-weight AI model, Inkling, Apache 2.0 license, no restrictions.

2025-08-21 · Source: Rohan's Bytes · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Intermediate, long

Summary

The July 16, 2026, AI intelligence brief details several key industry advancements. Mira Murati's Thinking Machines Lab released Inkling, a 975B-parameter (41B active) open-weights multimodal model with a 1M-token context window, ranking #10 in Frontend Code Arena and achieving 74.1% on MCP Atlas for agentic tasks. LMCache demonstrated up to a 10.7x speedup in LLM inference by reusing KV cache, supporting AMD MI300X. Meta's Muse Spark 1.1 entered the agentic coding race with competitive pricing (\$1.25/\$4.25 per 1M input/output tokens), outperforming Claude Opus 4.8 on some benchmarks. OpenAI revealed the prompt for GPT-5.6 Sol Ultra's math proof, utilizing 64 subagents and adversarial agents, while GPT 5.6 Terra shows poor intelligence-cost efficiency. Goldman Sachs projects a 24-fold increase in AI agent token use by 2030. Databricks' tests found GLM-5.2 rivals top closed models like Claude Opus 4.8 in enterprise coding at \$1.28/task, significantly cheaper than Opus 4.8's \$1.94/task. Apple also sued OpenAI, alleging hardware secret theft.

Key takeaway

For AI Engineers and Directors of AI/ML optimizing LLM deployments, you should prioritize evaluating open-weight models like Inkling and GLM-5.2, which now rival closed-source performance at significantly lower costs. Implement inference acceleration solutions such as LMCache to achieve up to 10.7x speedups, especially for agentic workloads where token usage is projected to multiply 24 times by 2030. This strategic shift can drastically reduce operational expenses while maintaining or improving model quality, making cost-efficiency a critical competitive advantage.

Key insights

Agentic AI models and efficient inference techniques are rapidly advancing, driving both performance and cost-effectiveness.

Principles

• Open-weight models can rival closed-source performance.
• KV cache reuse significantly boosts LLM inference.
• Agentic workloads demand cost-optimized orchestration.

Method

LMCache manages KV cache across memory tiers to reuse attention states from long prompts, reducing recomputation for repeated or overlapping text in agentic workflows.

In practice

• Deploy LMCache for agentic LLM inference.
• Evaluate open models like Inkling and GLM-5.2.
• Utilize adversarial subagents for complex problem-solving.

Topics

• Open-weight LLMs
• LLM Inference Optimization
• AI Agent Orchestration
• Multimodal AI
• AI Benchmarking
• AI Cost Efficiency
• AI Legal Disputes

Code references

• LMCache/LMCache

Best for: CTO, VP of Engineering/Data, MLOps Engineer, AI Scientist, Director of AI/ML, AI Engineer

Related on AIssential

• Ex-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China — Inkling leads U.S. open-weights models in agentic tasks and efficiency but struggles with factual accuracy and cost.
• [AINews] not much happened today — The AI frontier is rapidly expanding with competitive open-weight models, shifting focus to efficiency, specialized architectures, and robust inference marketplaces.
• GLM-5.2 is probably the most powerful text-only open weights LLM — GLM-5.2 is a powerful, open-weights LLM with a vast context window, leading benchmarks but consuming more tokens.
• Your LLM Isn’t Thinking — It’s an Engineer Pulling Weights at 10,000 Tokens Per Second — Efficient LLM inference engines, particularly vLLM, are critical for scalable, cost-effective AI serving by optimizing GPU memory and compute utilization.
• not much happened today — The AI landscape is rapidly evolving towards specialized, agentic systems, with open-weight models gaining significant ground against proprietary solutions.
• GLM 5.2 and the coming AI margin collapse (part 1) — Open-weights models like GLM 5.2 are collapsing AI inference margins by offering comparable quality at significantly lower costs.

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 Rohan's Bytes.