AIME

GitHub - SakanaAI/text-to-lora: Hypernetworks that adapt LLMs for specific benchmark tasks using only textual task description as the input Hypernetworks that adapt LLMs for specific benchmark tasks using only textual task description as the input - SakanaAI/text-to-lora

SakanaAI's new Text-to-LoRA (T2L) uses a hypernetwork to generate task-specific LoRAs from simple text descriptions—no expensive fine-tuning required.

✅ Compresses 100s of adapters
✅ Generalizes to unseen tasks
✅ ICML 2025 Paper & Code: github.com/SakanaAI/tex...

GitHub - jamiepine/voicebox: The open-source voice synthesis studio powered by Qwen3-TTS. The open-source voice synthesis studio powered by Qwen3-TTS. - jamiepine/voicebox

Voicebox: open-source, locally run TTS studio—no cloud, no subscriptions.
✅ Powered by Qwen3-TTS for expressive voice cloning
✅ Multi-track editor + inline audio editing
✅ Tauri/Rust app: 10× smaller than Electron
✅ MIT license, full privacy

github.com/jamiepine/vo...

Home - AKI.IO Token-based access to leading open-source AI models on EU infrastructure. Evaluate, build and scale your AI product without self-hosting or vendor lock-in.

AKI.IO is now live: Curated open-source and open-weight models such as #MiniMax M2.5, #Apertus 70B, #Qwen Image Edit and many more as an API – hosted entirely in European data centers w/o hyperscalers. Happy to get your feedback!
The playground is open, API key via free registration at aki.io

GitHub - QwenLM/Qwen3.5: Qwen3.5 is the large language model series developed by Qwen team, Alibaba Cloud. Qwen3.5 is the large language model series developed by Qwen team, Alibaba Cloud. - QwenLM/Qwen3.5

Qwen3.5 is out: Alibaba's open-weight series built for agentic AI with native multimodality.
✅ Flagship: 397B total / 17B active params (MoE)
✅ 1M-token context → 2h audio/video in one pass
✅ 60% cheaper, 8× more efficient than predecessor
✅ MIT license, full open weights

github.com/QwenLM/Qwen3.5

microsoft/VibeVoice-ASR · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Microsoft's VibeVoice-ASR transcribes 60-minute audio in a single pass—no chunking needed.
✅ 9B params, 64K-token context
✅ ASR + speaker diarization + timestamps in one inference
✅ MIT license, fully open source
A leap for meeting/podcast transcription 👇
huggingface.co/microsoft/Vi...

MiniMax M2.5: 更快更强更智能，为真实世界生产力而生

MiniMax M2.5 is out: a frontier model optimized via massive RL for agentic workflows.

Forge RL Framework enables near-linear scaling across 10k+ real agent scenarios.
Weights not yet public—MiniMax historically releases them later.
www.minimax.io/news/minimax...

GitHub - zai-org/GLM-5: GLM-5: From Vibe Coding to Agentic Engineering GLM-5: From Vibe Coding to Agentic Engineering. Contribute to zai-org/GLM-5 development by creating an account on GitHub.

GLM-5 is Z.ai's new MoE flagship (744B total/40B active) built for agentic engineering.
✅ #1 open-source on Vending Bench 2
✅ Closes gap with Claude Opus on CC-Bench-V2
✅ DeepSeek Sparse Attention for efficient 200K context
✅ Apache 2.0 license, commercial use allowed

github.com/zai-org/GLM-5

GitHub - ace-step/ACE-Step-1.5: The most powerful local music generation model that outperforms most commercial alternatives, supporting Mac, AMD, Intel, and CUDA devices. The most powerful local music generation model that outperforms most commercial alternatives, supporting Mac, AMD, Intel, and CUDA devices. - ace-step/ACE-Step-1.5

ACE-Step v1.5 is out: an open-source music generation model that runs locally with <4 GB VRAM.
8 diffusion steps → full songs in ~2s (A100)
4-min tracks with lyrics, 50+ languages
MIT license, full training code included
A leap for accessible, commercial-grade audio AI 👇
github.com/ace-step/ACE...

Qwen/Qwen3-Coder-Next · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Qwen3-Coder-Next is out: an open-weight MoE model (80B total / 3B active params) built for agentic coding workflows.
✅ 256K context length
✅ Tool-use & multi-step reasoning optimized
✅ Apache 2.0 license for local/dev use
Great step for open coding agents 👇
huggingface.co/Qwen/Qwen3-C...

GitHub - QwenLM/Qwen3-ASR: Qwen3-ASR is an open-source series of ASR models developed by the Qwen team at Alibaba Cloud, supporting stable multilingual speech/music/song recognition, language detectio... Qwen3-ASR is an open-source series of ASR models developed by the Qwen team at Alibaba Cloud, supporting stable multilingual speech/music/song recognition, language detection and timestamp predicti...

Qwen3-ASR drops: open-source speech recognition that transcribes speech, music, and singing across 52 languages — with accuracy rivaling GPT-4o and Gemini.

1.7B & 0.6B variants. Unified streaming/offline inference. Apache 2.0.

github.com/QwenLM/Qwen3...

GitHub - deepseek-ai/DeepSeek-OCR-2: Visual Causal Flow Visual Causal Flow. Contribute to deepseek-ai/DeepSeek-OCR-2 development by creating an account on GitHub.

DeepSeek OCR 2 is a 3B VLM that reads documents like humans do. "Visual Causal Flow" dynamically reorders tokens by semantic meaning, not left-to-right, unlocking 91.09% accuracy on complex layouts.

Fully open source (Apache 2.0).

github.com/deepseek-ai/...

Kimi released K2.5 — a native multimodal model trained on 15T visual-text tokens that generates full interactive UIs from prompts and orchestrates 100-agent swarms for complex tasks. 4.5× faster execution, 59% productivity boost. Open weights available now.
www.kimi.com/blog/kimi-k2...

GitHub - QwenLM/Qwen3-TTS: Qwen3-TTS is an open-source series of TTS models developed by the Qwen team at Alibaba Cloud, supporting stable, expressive, and streaming speech generation, free-form voice... Qwen3-TTS is an open-source series of TTS models developed by the Qwen team at Alibaba Cloud, supporting stable, expressive, and streaming speech generation, free-form voice design, and vivid voice...

Alibaba released Qwen3-TTS, a new text-to-speech model with discrete multi-codebook LM architecture under Apache license. Features 97ms synthesis latency, 3-second voice cloning, and 10-language support including German. Available on Hugging Face and ModelScope.

github.com/QwenLM/Qwen3...

NVIDIA just dropped PersonaPlex - a speech-to-speech model that lets you control AI personas through text prompts AND voice conditioning!

github.com/NVIDIA/perso...

zai-org/GLM-4.7-Flash · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Z.AI just released GLM-4.7-Flash - a 30B-A3B MoE model that dominates the 30B parameter class!

Perfect balance of power & efficiency for enterprise deployment. Supports vLLM, SGLang & native tool integration.

huggingface.co/zai-org/GLM-...

black-forest-labs/FLUX.2-klein-base-9B · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

FLUX.2 [klein] 9B Base released by @BlackForestLabs

🔥 9B-parameter undistilled foundation model
⚡ End-to-end inference in <1 second
💻 Runs on RTX 4090+ (~29GB VRAM)
🎨 Perfect for fine-tuning & research

Non-commercial license only

huggingface.co/black-forest...

GitHub - zai-org/GLM-Image: GLM-Image: Auto-regressive for Dense-knowledge and High-fidelity Image Generation. GLM-Image: Auto-regressive for Dense-knowledge and High-fidelity Image Generation. - zai-org/GLM-Image

Z.AI released GLM-Image, an innovative image generation model that establishes new benchmarks in specific application areas through its hybrid architecture.

github.com/zai-org/GLM-...

Alibaba Cloud just dropped Qwen3-VL-Embedding—a powerful, open-source multimodal embedding model that unifies text & images in one semantic space.

✅ SOTA on MTEB & MM-Retrieval
✅ Great for RAG, multimodal search
✅ Compact & production-ready

🔗 qwen.ai/blog?id=qwen...

Small Language Models for Efficient Agentic Tool Calling: Outperforming Large Models with Targeted Fine-tuning As organizations scale adoption of generative AI, model cost optimization and operational efficiency have emerged as critical factors determining sustainability and accessibility. While Large Language...

Researchers at Amazon Web Services (AWS) have published the paper "Small Language Models for Efficient Agentic Tool Calling: Outperforming Large Models with Targeted Fine-tuning."
arxiv.org/abs/2512.15943

LGAI-EXAONE/K-EXAONE-236B-A23B · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Korean LG AI Research releases K-EXAONE-236B-A23B - 236B-parameter MoE LLM with 23B active parameters.

💡 Hybrid Attention architecture, advanced reasoning & long-context processing
🌍 6-language support
⚡ 256K context window with smart memory optimization

huggingface.co/LGAI-EXAONE/...

NC-AI-consortium-VAETKI/VAETKI · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

NC AI consortium releases VAETKI - a collaborative 112.2B parameter MoE LLM built by 13 institutions!

💡 10.1B active parameters during inference
🧠 Part of broader initiative including VAETKI-VL-7B-A1B vision-language model

huggingface.co/NC-AI-consor...

upstage/Solar-Open-100B · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Upstage releases Solar Open - their flagship 102B-parameter MoE LLM trained from scratch!
💡 102B total / 12B active parameters
🧠 Pre-trained on 19.7T tokens for broad knowledge & robust reasoning
🔍 Built for transparency & customization under Solar-Apache License 2.0

huggingface.co/upstage/Sola...

skt/A.X-K1 · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

SK Telecom just released A.X K1 - Korea's FIRST hyperscale AI model with 519B parameters (33B active)!

It can be served with SGLang and vLLM.

huggingface.co/skt/A.X-K1

MiniMax has released MiniMax-M2.1, an open-source model designed to democratize top-tier agentic capabilities.

The model supports popular inference frameworks including SGLang, vLLM, and Transformers.

lnkd.in/gWwEbV9q

GitHub - QwenLM/Qwen-Image-Layered: Qwen-Image-Layered: Layered Decomposition for Inherent Editablity Qwen-Image-Layered: Layered Decomposition for Inherent Editablity - QwenLM/Qwen-Image-Layered

Qwen has released Qwen-Image-Layered, an innovative model capable of decomposing images into multiple RGBA layers, under the Apache 2.0 license.

github.com/QwenLM/Qwen-...

Meta Seal - State-of-the-Art Open Source AI Watermarking Open-source framework for invisible, robust watermarking across audio, image, video, and text. State-of-the-art AI watermarking from Meta's FAIR team.

Meta released Seal, a watermarking model, to advance the state of content authenticity and attribution.

The framework provides invisible, robust watermarking capabilities across all modalities including audio, image, video, and text.

facebookresearch.github.io/meta-seal/

GitHub - zai-org/GLM-4.5: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models - zai-org/GLM-4.5

Z.AI released GLM-4.7, their latest large language model with significant advancements in coding capabilities.

The model demonstrates competitive performance across 17 benchmarks incl. reasoning, coding, & agent tasks, with notable improvements over previous versions.

github.com/zai-org/GLM-...

GitHub - deepseek-ai/DeepSeek-V3.2-Exp Contribute to deepseek-ai/DeepSeek-V3.2-Exp development by creating an account on GitHub.

DeepSeek released V3.2, a model that harmonizes high computational efficiency with superior reasoning & agent performance. It surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro.

🥇 Also it achieved Gold-medal performance in the 2025 IMO and IOI

github.com/deepseek-ai/...

GitHub - Tongyi-MAI/Z-Image Contribute to Tongyi-MAI/Z-Image development by creating an account on GitHub.

Alibaba releases a powerful image generator, Z-Image, with 6B parameters in three variants that shows highly competitive performance against other leading models, while achieving state-of-the-art results among open-source models.

github.com/Tongyi-MAI/Z...

GitHub - PaddlePaddle/ERNIE: The official repository for ERNIE 4.5 and ERNIEKit – its industrial-grade development toolkit based on PaddlePaddle. The official repository for ERNIE 4.5 and ERNIEKit – its industrial-grade development toolkit based on PaddlePaddle. - PaddlePaddle/ERNIE

🚀 Baidu just released **ERNIE-4.5-VL-28B-A3B-Thinking** — open-source (Apache 2.0)!

✅ 3B active params
✅ 100% multimodal reasoning
✅ Visual reasoning, STEM, video understanding & “Thinking with Images”
✅ Tool use, precise grounding, dynamic zoom & search

👉 github.com/PaddlePaddle...