A 2B Model Just Beat Several 4B Models. Bigger Isn't Always Better.

Morning, {{first name | folks}}! Small models are getting surprisingly good, Shopify found a way to make one AI workload 27x cheaper, an Australian AI infrastructure company just landed OpenAI as a customer, Microsoft built an AI to attack AI, and Gemini learned when to look away. Let's get into it.

OpenBMB released MiniCPM5-2B, a 2-billion-parameter model built for phones and local devices, and it's outscoring several 4-billion-parameter models on average benchmark performance, 53.9 versus 51.1 for the best 4B competitor, despite being half the size. It's fully open too, Apache 2.0 license, training data included, and it's already been ported to run on 9 different AI chip makers, Nvidia, Ascend, Kunlunxin, and more.

This is the same story you're seeing everywhere lately, just at a different scale. Shopify proved a tiny, fine-tuned model can beat a frontier one on a narrow task. This is the open-source version of that same idea, a small model, trained well enough, closing a gap that used to only get closed by adding parameters. Efficient training is starting to matter as much as raw size.

Shopify CEO Tobi Lütke shared an internal experiment where a tiny 0.8B parameter model actually beat GPT-5.6 Sol at a narrow task, generating buyer profiles. They shrank the system prompt from 9,100 tokens down to 1,100, and throughput went from 2 million profiles a day to 72 million. Their GraphQL agent handles up to 2,000 requests a minute, running that on frontier models would run about $27 million a year. The fine-tuned model does it for roughly $1 million.

This doesn't mean small models are generally better, it means once a task is narrow enough, the frontier model's real job becomes teaching a cheaper one to do it, not doing it forever itself. Cost and speed stop being an afterthought at that point. They become the whole product.

Firmus signed OpenAI as an anchor customer for two new AI compute sites in Malaysia, pushing its total contracted capacity past 900 megawatts across all customers. The company runs 7 AI factories across Australia, Singapore, Indonesia, and now Malaysia, all built around Nvidia's next-gen Vera Rubin systems and Firmus's own prefabricated liquid-cooling platform.

Firmus's co-CEO put the bigger picture plainly, this is the moment "Asia-Pacific becomes a producer of intelligence, not just a consumer of it." Another sign the compute race isn't just a US-China story, infrastructure money is now flowing hard into Southeast Asia and Australia too.

Microsoft published its third annual Responsible AI Transparency Report, and the real headline is what agentic AI forced them to build. A new AI Red Teaming Agent whose entire job is probing Microsoft's other AI systems for vulnerabilities, plus an External Red Team Alliance pulling in 18 universities across 6 continents to stress-test from the outside too. Microsoft 365 Copilot, Foundry, and GitHub Copilot all now hold ISO 42001 certification, the actual international standard for managing AI systems.

The reasoning is the same one you've seen from other labs this month, once AI can act on its own across multiple steps, checking it once before launch isn't enough anymore. Microsoft reworked its entire safety standard around the tech stack itself, models, platforms, and apps as three separate risk categories, not one blanket policy.

Google is changing how Gemini handles video. Instead of treating every frame equally, it can decide where to look more closely and pull from frames, audio, or transcripts depending on what it needs to answer.

That sounds like a small technical change, but it matters. Google says the approach can cut token usage by up to 88% and analysis costs by up to 66%, while improving accuracy by up to 7%. Long videos are expensive to process, so teaching AI when not to look could be just as important as teaching it what to see.

Other AI Signals: