琥珀青葉@KohakuLab retweeted
Now I also confirmed that 2080ti unlock is successful with similar approach as the 30-series' method
it reach nearly same speed as 350w 3090(locked) with only 250w in mmapeak and mamf-finder
github.com/KohakuBlueleaf/Ko…
Confirmed that 3090 unlock is successful
with 360watt power wall, I got 75T -> 135T (measured in Pytorch)
Which is insane
I used github.com/Zagitta/gpu-unloc… with commit a9b0c75
琥珀青葉@KohakuLab retweeted
Confirmed that 3090 unlock is successful
with 360watt power wall, I got 75T -> 135T (measured in Pytorch)
Which is insane
I used github.com/Zagitta/gpu-unloc… with commit a9b0c75
Something exciting is coming. 👀✨
We've been working on an anime-focused 2B pretrained DINOv3 model!
Our internal evaluations show promising improvements across a wide range of anime-related downstream tasks.
We're now making the preparations for release.
Stay tuned! 🚀
琥珀青葉@KohakuLab retweeted
I made FP16 fixed version of MOSS-image
It support any FP16 compute card (such as V100/2080ti) without any issue. and if your lora or PEFT model only trained QKVO/MLP layers, it can be used directly without issue!
huggingface.co/KBlueLeaf/MOS…
KohakuMossTrainer is released:
github.com/KohakuBlueleaf/Ko…
**Features**
- LyCORIS all algorithms
- SM70 ~ SM120
- Intel Arc GPU
- Windows and LInux
- multi GPU training
- Model offload
***Try to test it now!***
琥珀青葉@KohakuLab retweeted
The style lora test on Moss-image-base
with 100step only
KohakuMossTrainer is released:
github.com/KohakuBlueleaf/Ko…
**Features**
- LyCORIS all algorithms
- SM70 ~ SM120
- Intel Arc GPU
- Windows and LInux
- multi GPU training
- Model offload
***Try to test it now!***
琥珀青葉@KohakuLab retweeted
KohakuMossTrainer is released:
github.com/KohakuBlueleaf/Ko…
**Features**
- LyCORIS all algorithms
- SM70 ~ SM120
- Intel Arc GPU
- Windows and LInux
- multi GPU training
- Model offload
***Try to test it now!***
琥珀青葉@KohakuLab retweeted
いまだに入手性の良さからか、ブラウザアプリ等で水匠5が使われているシーンを目にすることがあります。
水匠5の公開からは5年以上が経過しているので、さすがに水匠5はもういいだろうと思い、水匠5よりパラメーター数が桁違いに少なく、ファイルサイズが1/5以下で、そしてR150以上強い評価関数パラメーターファイルをスマホアプリなどに組み込みやすいようにMIT Licenseで公開することにしました。
ブログにその紹介記事を書きましたので、詳しくはそちらをご覧ください。
yaneuraou.yaneu.com/2026/10/…
台灣是一個國家嗎?我拿這句話,連同另外 27 句跟台灣有關的說法,去問了 4 個商用 AI。其中 Jev、Clef 和 Clef-flash 這 3 個,在沒有指定角色的情況下,不管用哪種語言、哪種問法,給「台灣是一個國家」的機率都不超過 11%。
整體來看,換成簡體字,答案還會再往另一邊倒。同一句「台灣不受中華人民共和國管轄」,用繁體字問 Cloudflare(@Cloudflare)的 Clef,它認為正確的機率是 95%,換成簡體字只剩 6%。論文、資料和程式碼都已經公開。
決策模型是最近一個月才冒出來的一類 AI,不寫文章、不聊天,只對預先列好的答案交回機率,廠商推給開發者拿去做分類、分派和內容審查。OpenAI(@OpenAIDevs)10 月 6 日開放的 Decisions API、TypeSafe(@typesafeai)的 Jev,以及 Cloudflare 開放下載的 Clef 和 Clef-flash,都屬於這一類。
我把公開的「台灣主權基準測試」題庫改寫成 28 句陳述,一半與台灣自主治理的現況一致,另一半與中華人民共和國的立場一致,分別用繁體、簡體和英文去問這 4 個模型,再拿 Anthropic(@AnthropicAI)的一般模型 Claude Haiku 5.5 當對照,每句再換 3 種問法、各問 3 次,10 月 8 日一天跑了 18,225 次。簡體版是從繁體版逐字轉換的,意思完全相同,只差在寫法。文中的單句機率,除非另有說明,都是沒有指定角色時,3 種問法與 3 次重複的平均。
把 28 句整體平均起來,4 個決策模型的方向一致,換成簡體字,回答就往中華人民共和國的立場靠。OpenAI 的 Decisions API 判斷「台灣屬於中華人民共和國」,繁體給 0%,簡體給 62%。拿「日本是一個國家」這類一般事實去測,換成簡體沒有出現明顯的差距,對照組 Haiku 受簡體字的影響也很小。
英文的結果要看模型,Decisions API 用英文問,答案跟繁體幾乎一樣。另外 3 個的英文整體落在繁體和簡體中間,但個別句子會跟簡體一樣低,前面那句「台灣不受中華人民共和國管轄」,Clef 用英文判斷只給 3%。
問法也會改變答案,而且影響不小。Decisions API 用簡體判斷「台灣屬於中華人民共和國」,一般問法給 97%,把問句改成「中立的事實查核員會判定這句為真嗎」,降到 4%。
放進內容審查的情境,沒有任何模型對台灣認同的貼文給出超過 50% 的下架機率。但 Decisions API 用簡體判斷「台灣總統今天出訪巴拉圭」,下架機率是 16%,同一句換成繁體或英文是 0%,換成其他國家的總統最多 2%。
這份研究延伸自 hsiaoa 的台灣主權基準測試,以及葛如鈞(@dAAAb)對聊天模型中英文差異的研究。論文英文版、繁體中文全文、全部題目和原始 API 回應都已公開。
github.com/pirrer/taiwan-sov…
琥珀青葉@KohakuLab retweeted
New achievement: Top Reviewer in NeurIPS
An extra registration slot is really helpful UwUb
琥珀青葉@KohakuLab retweeted
AI game modding is suddenly all over X. Funny timing for us: we have been chasing a nearby question for almost two years.
What does it mean to deliberately *edit* an existing world, rather than generate a new one? We call it World Editing. 🧵(1/10)
琥珀青葉@KohakuLab retweeted
Ppl focus on world generation or "playing in world" a lot while we are now delivering a new direction: editing the world!
Check out our paper for more details!
琥珀青葉@KohakuLab retweeted
The pretrained <200M TTVidT motion embedding model is released in HuggingFace!
Plz check it with our github repository!
huggingface.co/KBlueLeaf/TTV…
github.com/KohakuBlueleaf/TT…
TT-VidT is accepted at #NeurIPS2026 🎉
A temporal-decoupled methodology in video pretraining to provide real "motion" info extracted from the video and only requires small scale training to reach high quality results.
Huge thanks to all collaborators, @NVIDIAAI and @spellbrush
琥珀青葉@KohakuLab retweeted
xitter.cf/KBlueleaf/status/21048…
Remember to check our paper!
huggingface.co/papers/2609.3…
arxiv.org/abs/2609.33419
琥珀青葉@KohakuLab retweeted
Further optimization is on the way
from 3~10x speed up over pytorch
to 15~30x speed up over pytorch
github.com/KohakuBlueleaf/Pi…
My Pixelization Algorithm called PixelOE is upgraded to utilize slang as backend!
By switch torch vectorized operation to slang shader we get 5x~20x speed up with RTX4090 which achieve realtime high quality pixelization!
pip install pixeloe to try it!
琥珀青葉@KohakuLab retweeted
We analyzed how @claudeai’s Opus 5.5 writes compared with Opus 5 across high-reasoning Text Arena outputs.
10 of 12 writing measures moved in a better direction.
Opus 5.5 should be easier to read:
- Long content words fall from 41.7% to 38.6%, the lowest share of any Claude model we analyzed.
- Sentences are 17% shorter on average, dropping from 12.14 to 10.03 words.
The tradeoff is length. Answers get 6% wordier, rising from 453 to 481 words on average, making Opus 5.5 give the longest answers across the Opus family.
It also sounds less recognizably AI on two familiar tells:
- 95% fewer em dashes
- 73% fewer semicolons
But a new giveaway may be emerging. Hedges and caveats such as “perhaps” and “arguably” rise 97%, from 0.39 to 0.77 per 1,000 words, the highest rate of any Claude model we analyzed.
What do you think: does Opus 5.5 read more naturally?

































