markus

[markus @duckllm]

@real_learae A third party could reuse MiniMax’s tokenizer, but wouldn’t normally reproduce its internal response-ID clock. Since OpenCode also issues a separate log ID, the matching clock likely comes from upstream. That points to MiniMax infrastructure and not just MiniMax-derived weights.

[markus @duckllm]

@real_learae Tested this across 34 alternating M3/Space Bunny responses. The first 8 hex chars decode to request time using a 2023-01-01 epoch. M3 body ID = MiniMax trace-id; OpenCode assigns Bunny a separate log UUID. Strong MiniMax serving-plane evidence.

[markus @duckllm]

@RyanLeeMiniMax @MiniMaxAgent Amazing work 🚀

[markus @duckllm]

@opencode @ivanfioravanti Any leads?

[markus @duckllm]

@ivanfioravanti @Youssofal_ Safe travels! I'm super excited to see what you'll report from your Minimax visit 🤩

[markus @duckllm]

@xiaogaifun 这个做��太棒了！想请教一下，女儿那些原始画作是怎么数字化的？是用扫描仪，还是手机/相机拍的？后期有没有做光线、透视或纸张背景校正?

[markus @duckllm]

@Thom_Wolf 🤗🐥🚀

[markus @duckllm]

@ivanfioravanti New Minimax model ❤️

[markus @duckllm]

@taohe669360 💯

[markus @duckllm]

@ZeyuanAllenZhu @syhw Thank you for all the amazing research!

[markus @duckllm]

@brianchew @MiniMax_AI @olive_jy_song @Thom_Wolf @huggingface https://t.co/bzj17BHue5

[markus @duckllm]

@jaminball @palak_go @simon_mo_ Great podcast! I'm also super happy that @simon_mo_ spoke more broadly about Chinese labs. Very insightful and nuanced!

[markus @duckllm]

@dyfan22 @maksym_andr Any plans to add Minimax M3?

[markus @duckllm]

@ivanfioravanti Have you tried using GLM 5.2 for planning and M3 for execution?

[markus @duckllm]

@tomgreenwald @FredaDuan

[markus @duckllm]

@tomgreenwald Very cool! Have you also looked into Minimax M3 for execution?

[markus @duckllm]

@SakanaAILabs @hardmaru @FredaDuan

[markus @duckllm]

@FredaDuan GLM 5.2 looks amazing... but M3 also seems quite good and a fair bit cheaper

[markus @duckllm]

@ArtificialAnlys So GLM-5.2 for planning and M3 for execution?

[markus @duckllm]

@elliotarledge @ttanvali Yes, it would be great to see if pi or opencode change the results for open models.