beichuan

Building with AI | Coding & Agents

[beichuan @beichuan96]

据 Artificial Analysis 的独立测速，GPT-6.1 Sol (max) 平均大约 66 token/s，上一代 GPT-6 Sol (max) 还有 76 token/s。

跑了一晚上，体感确实慢了不少。

这突然让我想起 2026 年 6 月 8 日参加小米 MiMo-V2.5-Pro-UltraSpeed 内测的时候。

当时测试能跑到 1000 token/s 左右，最高甚至见过 1300 token/s。

那个速度真的快得有点离谱。虽然模型效果确实还差一些，但单说交互体验，是真的舒服。

当然，两者的测试方式、模型能力、推理过程都完全不同，不能直接拿速度硬比。

但还是忍不住幻想：

什么时候 OpenAI 能在保持这种在线思考能力的同时，把输出速度也干到 200+ token/s？

那体验应该真的会完全不一样。
https://t.co/I44M5YOlNv

[beichuan @beichuan96]

@chasen_liao 我们都感觉有点慢，这是真的😂

[beichuan @beichuan96]

@iecbearli @chasen_liao 可以更新下codex 和 cli就能看到了

[beichuan @beichuan96]

@Saccc_c @lemomo_ai 有种当年小龙虾又回来的感觉，尽管是换了大脑。

[beichuan @beichuan96]

@Arvin010717 @Saccc_c @thsottiaux 这次貌似需要手动检查更新，我是早上发现手动更新的

[beichuan @beichuan96]

@thsottiaux Aww, Plus users aren’t eligible to use it. And Musk already bought the domain https://t.co/Ya5tvpU6bJ. This is so awkward.

[beichuan @beichuan96]

@thsottiaux the english version is here.
It’s a pretty literal translation, so bear with it.

[beichuan @beichuan96]

@thsottiaux Is Tibo going to end up like this too?

[beichuan @beichuan96]

@thsottiaux Bro, putting all the nonsense aside once again, after DevDay, will it be able to surpass Opus 5.5 overall? If so, then I can go Pro.

But, but... I’m really afraid I haven’t woken up yet.

[beichuan @beichuan96]

@Hywangjfwang @thsottiaux 能看懂，之前发中文回复过😭

[beichuan @beichuan96]

@rico10000000 @thsottiaux 等我忙完去改一个

[beichuan @beichuan96]

@ClinsonJam2375 @thsottiaux 哈哈😄

[beichuan @beichuan96]

@leikr15 @thsottiaux A不说废话蒙头干事，就是封号太狠😭

[beichuan @beichuan96]

@HarriesSteele @iamhyhor 这就必须点赞👍了呀

[beichuan @beichuan96]

@CureVitanr @thsottiaux 😭万万没想到这波解释对标梁哥了

[beichuan @beichuan96]

@thsottiaux Also, there’s another piece of news: Is the 5-hour limit for Plus going away?🤩

[beichuan @beichuan96]

this battle of wits between these genius sellers is way too funny.
it cracks me up.

[beichuan @beichuan96]

@ashfold 已经认不清Tib在干嘛了，如果没有不断的reset，估计很多人退订了。
但是A÷怎么能稳定不被封号才是关键😂

[beichuan @beichuan96]

@thsottiaux Seeing you hesitate to say something, I’m really curious.

[beichuan @beichuan96]

When I applied for the MiMo Desktop beta, I got approved within just a few hours. Since then, I’ve been spending about an hour and a half testing it every night.

During testing, I ran into a few pretty obvious issues:

1.Long-running tasks often get stuck, time out, or get interrupted. After I reported this, the team reset usage quota.

2.The same tools sometimes get called repeatedly. For the same task, it takes noticeably longer than clients like Codex, A, and DS.

3. It also seems like the context doesn’t automatically get compressed after usage reaches a certain threshold. So when a single task or conversation runs for too long, it tends to become laggy or get stuck.

[Xiaomi MiMo Developers @XiaomiMiMoDevs]

🛠️ MiMo-V2.6 update: tool-call repetition, diagnosed & fixed.
After the MiMo-V2.6 series models launched, we noticed them sometimes repeating identical or highly similar tool calls — burning context and stalling tasks, especially in MiMo Desktop, MiMo Code and OpenCode.
Root cause
A "reward blind spot" in scaling RL: when rewards only track final-answer correctness, inefficient behaviors along the way go unnoticed — and get amplified as training scales. Concretely, our flooding penalty only kicked in at >32 tool calls per turn, so anything below that threshold went completely unpunished.
The fix
We trained a light-weight repetition-specialized RL teacher — just 12 steps, ~7k examples — and merged it into the main model via MOPD, at roughly 4% of the cost of a full mixRL retrain. Repetition dropped sharply across harnesses and context lengths, while benchmarks held steady.

📦 Updated models, open-sourced (MOPD suffix): https://t.co/u81xJigdkn
📝 Full postmortem: https://t.co/SfFokWC0iD

Huge thanks to our community for the patience and feedback 🙏 Updated models go live on our API platform Sep 25, 06:00 (UTC+8), names unchanged. And for MiMo Desktop users: everyone's remaining quota in the current window will be reset.