Local LLMを中心に、AI技術系の投稿とネタツイ Tesla v100 32GB×2枚と共に生きる。 最近はFlash Nextちゃんと遊んでいます アイコンは羊野郎さん@maton_3110
Qwen3.8 27Bから浮気中
Joined August 2019
- Tweets16.5K
- Following1.9K
- Followers7.4K
- Likes49.2K
Ryu retweeted
Introducing Fish Audio Realtime STT (preview): speech-to-text that understands the conversation, not just the words.
It knows when to answer, when to keep talking, and when to stop.
Ranked #1 of 20 on Cekura’s STT benchmark (2.21% overall word error rate).
Built for voice agents. Try it now 👇
Ryu retweeted
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Ryu retweeted
This is Genie 4, our most realistic and interactive world model yet. 🌐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
Ryu retweeted
Liquid AI's LFM3 is coming 💧
A Liquid AI engineer just opened a llama.cpp pull request adding the LFM3 family:
• LFM3-1.7B, plus a Base version
• LFM3-VL-2B, a vision model (the PR says video input comes later)
LFM3 mixes short convolution layers with sliding-window and full attention. The Hugging Face repos are still private.
Ryu retweeted
Announcing three major scientific discoveries, one each in mathematics (harmonic analysis), computer science (complexity theory), and physics (quantum field theory).
AI starting to help us improve our understanding of the universe! openai.com/index/building-sc…
Ryu retweeted
Kojo, our latest model, is out.
Chat with it for free, or build on it through an OpenAI-compatible API.
kojo.sh
Ryu retweeted
The Qwen4 model series is upcoming.
It features a new architecture:
QSA (Query-Sparse Attention) + GDN: Greatly reduces KV cache costs and boosts prefill speed for million-token context. Official data shows a 7.6× prefill speedup at 1M context length (per Tongyi Cloud).
Gated Residual: Four parallel information-flow channels to improve gradient flow during deep reasoning.
N-gram Embedding: Augmented with 51B embedding parameters. It adds almost no inference compute while expanding the model’s memory capacity.
Muon Optimizer: Deployed as a replacement for or hybrid with AdamW. It is adapted to the new architecture’s scaling laws and improves training convergence for large models.
Among these, Qwen4-Max is positioned to compete against Opus 5.5.
Anticipated lineup:
Ryu retweeted
早いのと、目がいいので自分で確認して修正ってループをSolよろしくしてくれるんだけど、しばらく使ってるといつものあれ（capacityエラー）が頻発するようになった。