Ox Alpha: a free, 1M-context stealth model topping coding charts
Anonymous on OpenRouter, it reportedly beats GPT-5.6 Sol and Claude on DeepSWE. Free ends ~Aug 27; prompts are logged, and fingerprints point to GLM-5.x.
Copy markdown
Free, and topping the coding charts
`stealth/ox-alpha` went live on OpenRouter with a 1,048,576-token context, 131K max output, and text/image/video input. One independent run clocked it at 80% on DeepSWE — reportedly ahead of Claude Fable 5 (65%) and GPT-5.6 Sol (52%) — but that's a single unofficial score, not a verified benchmark.
The catch: the free window closes ~Aug 27
Access is free during a one-week preview that opened Aug 20, and OpenCode Go is routing it free for six days from Aug 21; pricing after is 'TBD.' If you want to stress-test a frontier-class coding model on your own repo for nothing, this is the weekend to do it.
Read the data terms before you pipe secrets through it
OpenRouter warns that prompts and completions for Ox Alpha are retained by the anonymous provider — not used for training on this release, but logged. The OpenCode route claims zero data retention. Either way, keep customer PII and production secrets out of a model whose operator you can't name.
The fingerprints point to Zhipu's GLM-5.x
It's officially anonymous, but independent forensics — tokenizer matches, video-encoder behavior, error codes — reportedly line up with an unreleased GLM-5.x flagship from Zhipu, the lab behind last week's GLM-5.3. Treat the attribution as speculation until someone confirms it.
Elsewhere: Diffusers 0.40 adds local video, music and audio
Hugging Face's Diffusers 0.40.0 ships new video/music/audio pipelines with Wan-Animate 2, LTX-2.5 and MiniMax-H3 support, plus tensor-parallel inference. MiniMax-H3 renders a 5-second clip with synced audio in ~44s on a 48GB card — a genuine weekend unlock for local generative media.