しがないフリーランスエンジニア。zennのscrapに生息。競馬好き。
日本 神戸
Joined February 2013
- Tweets22.4K
- Following609
- Followers1.3K
- Likes14K
kun432🇯🇵 retweeted
Can a 27B model push the Pareto frontier of autonomous research and beat 1T–2.8T giants?
Meet AREX-2 (27B), spearheaded by BAAI in collaboration with USTC and RUC (built on Qwen3.8-27B). Turning Recursive Self-Improvement (RSI) into a practical test-time scaling paradigm, we pinpoint two keys to self-evolution: Deep Reflection & Long-Horizon Execution. By reinforcing algorithmic optimization tasks, AREX-2 emerges with exceptional deep reflection and long-horizon execution capabilities—significantly boosting its performance oncomplex coding and deep research to top MLE-bench Lite (81.8) and set new ≤40B SOTA on BrowseComp (84.0) & HLE (52.6)👇
kun432🇯🇵 retweeted
Your OCR says it’s done—but did it get everything right? 🤔 In high-stakes workflows, even a small error can be costly.
🚀 Meet the MiniCPM-OCR suite: an open-source solution to parse fast and guard the results.
Two new models:
🔹 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. Compact, accurate, and fast.
🔹 MiniCPM-OCR Jev — A 2B decision model that helps assess parsing accuracy, with performance approaching that of Qwen3.5-397B-A17B on JevOCRBench.
⚡ What’s new?
Small model, SOTA results: MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2.
A sharp goalkeeper 🛡️: MiniCPM-OCR Jev is model-agnostic, checking text, formulas, and tables from different OCR parsers. It achieves 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8).
Up to 128 pages per minute 🤯: With DSpark, MiniCPM-OCR runs at 1.49× the speed of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro. The SDK’s concurrency optimizations further boost throughput to as much as 128 pages per minute.
Deploy where you work: Day-0 support for Arm, Houmo AI, Intel, and NVIDIA, plus 10 more chip vendors through FlagOS.
Ready for your stack: vLLM, SGLang, llama.cpp, and Ollama, with quantized models, fine-tuning support, and Agent Skills.
🔗 GitHub: github.com/OpenBMB/MiniCPM-O…
🤗 Hugging Face:
MiniCPM-OCR 1.0：huggingface.co/openbmb/MiniC…
MiniCPM-OCR Jev：huggingface.co/openbmb/MiniC…
👉 Try the Demo：platform.modelbest.ai/consol…
Use the API：platform.modelbest.ai/consol…
kun432🇯🇵 retweeted
GPT-6.1 Terra beats Astra in our vision benchmarks
wtf OpenAI?
had to double-check the results
- text
- object detection
- counting
also ~47% faster and ~92% cheaper
↓ leaderboards and examples
kun432🇯🇵 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/ 🧵
kun432🇯🇵 retweeted
Full-duplex 2話者データの音声分離・復元モデル「DialogueSidon」の高速版「FastDialogueSidon」を公開しました！
1ステップ推論＋軽量デコーダで、DialogueSidon公式実装比で約100倍の高速化と同等の性能を実現しています（H100上で計測）
GitHub：github.com/higurashi-ai/Fast…
Huggingface：huggingface.co/higurashi-ai/…
kun432🇯🇵 retweeted
Voice agents are either fast or smart. ConvFill proposes a way to bridge that gap by training a smol talker model (0.6B!) to talk with cheat notes from a large reasoner model.
It works like this: after the user's speech is transcribed, the talker replies quickly to buy itself time while the reasoner thinks. The reasoner then generates an appropriate response, which the talker reads to continue the conversation.
It's a bit like if you went on a panel with rocket scientists and had a little earbud with your own expert that could tell you what to say. You just reply quickly so as not to blow your cover, and then you begin saying what the expert in your earbud is telling you.
Playing with the model, I think the main idea is there and this could be a great direction to go, but the talker model should be a bit smarter for it to work like magic. Still, really cool direction!
ConvFill was developed by students from the University of Washington UbiComp Lab. I found it cool, so I reached out to them so we could build a really good demo for it together. They are presenting their paper at EMNLP this week.
Go try the demo: huggingface.co/spaces/smolag…
kun432🇯🇵 retweeted
Project Maya v1.1.29 is here, and it's a big one: MiniMax-M3.1-Flash now runs on Maya.
That makes three frontier open models you can run at home on your own GPUs: GLM-5.3-Flash, DeepSeek-V4.1-Flash and now MiniMax-M3.1-Flash, a 428B model with 23B active per token and up to a 1M context.
MiniMax gets its own engine, built by @needmorevram. We brought it up to the same standard as GLM and DeepSeek before shipping it:
- Runs on NVIDIA from V100 up, on Linux and Windows
- Fits itself to your PC: the setup picks the context your memory can hold, and the engine plans VRAM and RAM instead of failing
- A compact cache by default, so long contexts take half the memory with no loss in quality
- Pictures, tool calls and every one of MiniMax's thinking levels, with its own recommended settings out of the box
- The vision tower stays off your GPU until you send a picture, so the VRAM goes to the model
Huge thanks to @needmorevram, and to everyone who tests, reports and contributes to Maya. This is what community open source looks like.
Next up (WiP): our own Maya GGUF quants of MiniMax-M3.1-Flash, small enough for a 128 GB PC.
Link in the reply 👇
kun432🇯🇵 retweeted
GPT-6.1 Terra: delivering fast and capable computer use at a lower price.
GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.
🤖 Made with AI
kun432🇯🇵 retweeted
MiniMax-M3.1-Flash can now run locally!
We made 2-bit GGUF retain 91.5% top 1 accuracy while being 45% smaller. Run on 150GB RAM.
MiniMax-M3.1-Flash is the strongest open model for its size & among the best overall.
GGUF: huggingface.co/unsloth/MiniM…
Guide: unsloth.ai/docs/models/minim…
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
kun432🇯🇵 retweeted
In the Bible, humans tried to build the Tower of Babel but failed when language split them. Today, AI Agents are trying to "build" tasks across 23 languages—and facing the exact same barrier. 🧱🗣
Qwen Team introduces BabelArena, a new open-source multilingual Agent benchmark on 🧵👇
🔗 Dataset: modelscope.ai/datasets/Qwen/…
📄 Paper: modelscope.ai/papers/2609.23…
🏆 No single model dominates all benchmarks.
⚠ Low-resource languages show distinct failure modes: high rates of tool-use and control-flow errors (not just answer quality).
📈 Agents consume up to ~2x more tokens in low-resource languages without more interaction rounds.
🗣 Language consistency drops in structured-output tasks, with switches overwhelmingly pointing to English.
kun432🇯🇵 retweeted
ひい、まじかあ、AWSの容量が無くなるって言う日が来るとは思わなかったなー。
> Amazonのアンディ・ジャシーCEOは、AIインフラの需要急増により、AWSのクラウド容量が2027年まで大部分売り切れ状態で、2028年分も相当量が予約済みだと明らかにしました。
AWSのクラウド容量、2027年まで大部分が売り切れ状態とAmazonCEOが発言 jp.investing.com/news/compan…
kun432🇯🇵 retweeted
In this semester, I have a class for graduate school "Speech and Audio Processing." Materials are fully open; enjoy it!
大学院講義「音声音響工学」のスライドをすべて公開中です．まだ途中ですが楽しんて下さいね！
github.com/takamichi-lab/spe…
kun432🇯🇵 retweeted
Nanbeige 4.3 33b uses an old llama architecture, and that makes inference very costly
At 256K tokens (BF16, 1 sequence):
2 (K+V) × 52 layers × 8 KV heads × 128 dims × 262,144 tokens × 2 bytes ≈ 55.83 GB.
Qwen3.8-Flash-Next: ~6.76 GB including indexer/recurrent state (no MTP). Nanbeige uses ~8.3× as much cache memory.
KV cache quantization will be necessary and still this will quickly saturate your memory bandwidth.
Maybe that's the price to pay for this kind of accuracy for such a small model
kun432🇯🇵 retweeted
Meet Dessa, our AI model trained from scratch on nothing but speech.
We believe speech is the natural interface for technology. So we built Dessa.
And we’re inviting AI safety researchers to explore Dessa.
Learn more: block.xyz/dessa
kun432🇯🇵 retweeted
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
kun432🇯🇵 retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
kun432🇯🇵 retweeted
Run a 763B-parameter model locally. DeepSeek-V4.1-Flash-1017 on a 12GB GPU.
Frontier quality with GSQ-RCO quants, directly supported in the Strata inference engine.
Further info below. Much more to come. Follow for updates.
kun432🇯🇵 retweeted
Available now: NVIDIA Nemotron 3.5 Bolt E4B 4GB ⚡️
Our newest small open model is built for local assistants, interactive characters and robots that can carry on a conversation and use tools to get things done.
Watch Reachy Mini chat and move, with Bolt handling conversation and NVIDIA Cosmos handling vision. Both models run locally on Jetson Orin Nano 8GB.
Learn more: research.nvidia.com/labs/nem…