An Outpouring of Open Models — But the US Evaluation Agency Says "The Gap with Frontier Is Widening"
This month has been extraordinarily busy for the open-source AI model scene. Major open models dropped one after another: DeepSeek V4, Google's Gemma 4, Moonshot AI's Kimi K2.6, Xiaomi's MiMo 2.5, and Zhipu AI's GLM-5.1, among others. Florian Brand of Interconnects described it as "an intense month of back-to-back flagship model releases."
However, when CAISI (Center for AI Standards and Innovation), the US AI standards body, evaluated DeepSeek V4 Pro, it concluded that open models still lag behind American frontier models — and that the gap is widening over time. A cool-headed assessment that stands in stark contrast to Japan's DeepSeek fever.
The progress of open models is real, but the picture of frontier models outpacing that progress is a key lens for reading how the competitive landscape will shape up going forward.
AI Is Inflating College Grades — A Side Effect Creeping into Education
According to Axios, the spread of AI has sent college grades into abnormal inflation. The cause: students using AI to complete assignments can now easily earn high scores that would have been out of reach before.
At first glance it looks like "student performance is improving," but in reality it means the grading system itself is ceasing to function. If grades can no longer serve as a signal of ability, employers will go looking for other signals — and the value of a degree itself may end up eroded.
The relationship between AI and education is usually framed around "AI literacy education" or "skills for using AI," but this reporting throws a spotlight on a more fundamental problem: the collapse of the grading system itself.
Microsoft Open-Sources the "Strava for AI Coding"
Microsoft has published "AI Engineering Coach" on GitHub. It's a tool for visualizing and analyzing how you use AI coding tools such as Copilot, Claude, and Codex, and Hacker News commenters dubbed it "Strava for AI coding."
There's a genuine practical need for developers to quantitatively understand their own AI usage patterns. Which tasks do they lean on AI for? What's their acceptance rate for AI suggestions? How does that correlate with code quality? Data like this can feed into AI adoption strategies not just for individuals but for teams and entire organizations.
With AI coding tools proliferating, the arrival of a tool that takes a meta-level look at how they're being used signals that AI development workflows are maturing.
A "Software" Approach to Cutting AI's Energy Consumption
An article from The New Stack draws attention to energy-efficiency gains in AI through streaming inference — an approach that cuts AI power consumption at the software level, without deploying any new hardware.
Specifically, instead of the conventional batch approach that processes inference requests all at once, requests are handled incrementally in a streaming fashion, which curbs peak resource consumption. Because the technique doesn't depend on hardware upgrade cycles, it could be an attractive option, especially for data centers where power infrastructure is the bottleneck.
AI's energy consumption is turning into a societal issue — xAI's 50 gas turbines being a case in point — but demand-side efficiency is also a crucial part of the solution.
MTP Benchmarks on AMD's Strix Halo: 27B Gets Faster, 35B Actually Slower
Benchmark results following llama.cpp's MTP (Multi-Token Prediction) support on AMD's integrated GPU, the Strix Halo, have been shared on Reddit. Tests using Qwen3.6 models produced interesting, contrasting results.
With the 27B model, MTP boosted generation speed by 111% (7.63 → 16.15 t/s) and cut total time by 11.5%. The 35B model, on the other hand, ended up slower overall, with MTP's overhead outweighing the gains.
The data shows that MTP isn't necessarily effective at every model size — a fresh reminder that the balance between model size and hardware performance matters. As local LLM usage on AMD GPUs grows, first-hand reports grounded in real measurements like these are invaluable to the community.
---
**Sources:**
[Latest open artifacts (#21): Open model bonanza! Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others](https://www.interconnects.ai/p/latest-open-artifacts-21-open-model) (Interconnects)
[AI Sends Grades into Overdrive](https://www.axios.com/2026/05/16/ai-grade-inflation-college-classes) (Axios)
[Show HN: Strava for AI coding – analytics on your Copilot/Claude/Codex usage](https://github.com/microsoft/AI-Engineering-Coach) (GitHub / Hacker News)
[The software fix that could shrink AI's energy bill without new hardware](https://thenewstack.io/streaming-ai-energy-efficiency/) (The New Stack)
[Strix Halo Llama.cpp MTP Benchmarks: 27B Gets Much Faster, 35B Is Mixed](https://www.reddit.com/r/LocalLLaMA/comments/1teypb8/strix_halo_llamacpp_mtp_benchmarks_27b_gets_much/) (Reddit LocalLLaMA)