🔄 [GitHub Releases] PygmalionAI/aphrodite-engine - v0.22.0
https://github.com/dphnAI/sonar/releases/tag/v0.22.0
https://github.com/dphnAI/sonar/releases/tag/v0.22.0
GitHub
Release v0.22.0 · dphnAI/sonar
What's Changed
feat: add native Metal support by @AlpinDale in #1668
chore: optimize metal backend performance by @AlpinDale in #1669
perf: optimize GDN performance on Metal by @AlpinDale in #...
feat: add native Metal support by @AlpinDale in #1668
chore: optimize metal backend performance by @AlpinDale in #1669
perf: optimize GDN performance on Metal by @AlpinDale in #...
🗓️ Weekly GitHub Activity
🦙 llama.cpp
└ Release: b9966 → b10068
└ 102 commits
- Added support for Hunyuan 3 (hy_v3) with MTP speculative decoding #25395 #25641
- Added support for Minimax2 Eagle3 speculative decoding 259ae1d
- Added support for BitNetForCausalLM GGUF conversion #25769
- Implemented GGML_OP_LIGHTNING_INDEXER for DeepSeek V3.2/V4 on CPU and CUDA #24231 #25545
- Added fused hyper-connection ops for DeepSeek V4 to reduce graph splits #25585 #25702
- Added CUDA Virtual Devices support and enabled CUDA graphs on Volta and Turing architectures #25228 #25749
- Added Flash Attention via oneDNN graph API for SYCL on Intel Battlemage #25222
- Optimized CUDA MoE gate/up activation quantization, improving prefill times on RTX 5090 and Blackwell #25441
- Added auto-download of DeepSeek-Flash and Eagle3 speculative decoding sidecars from Hugging Face #25811
- Server now supports CORS configuration options and accepts null sampling parameters to request defaults #25655 #25538
- Added KleidiAI SME2 f32 kernel and improved hardware-specific kernel dispatch #24414 #25478
- Fixed CUDA crash when querying memory on devices with no available memory #25157
- Fixed Tensor Parallel execution for Phi3, Bert, Plamo2/3, and ChatGLM #25536
- Fixed quantization crash on DeepSeek-V4 i32 routing tables #25787
🔗 All changes | Latest release
🎨 stable-diffusion.cpp
└ Release: master-775-b5d8120 → master-782-b290693
└ 7 commits
- Support for AnimateDiff SD 1.5 motion modules v2 and v3 #1784 with img2video capabilities via the --init-img parameter #1789
- Support for ADetailer #1785
- Support for PiD 1.5 #1790
- Configurable reference image processing for edit models #1780
- Fixed cross attention and output projection token protection for Anima LoRAs #1786
🔗 All changes | Latest release
🤗 Fresh models trending on HuggingFace:
thinkingmachines/Inkling ♡1060
OpenMOSS-Team/MOSS-VL-Realtime ♡76
ai-sage/GigaAM-Multilingual ♡56
nineninesix/diamond-1.0 ♡43
ai-sage/GigaChat3.1-Audio-10B-A1.8B ♡38
rzgar/Bernini-R-S2V ♡37
acvlab/ABot-World-0-5B-LF ♡29
fal/ideogram-v4-instant ♡27
InternScience/Agents-A1-4B ♡26
sensenova/SenseNova-U1-8B-MoT-Infographic-V3 ♡26
OpenMOSS-Team/MOSS-VL-Instruct-0708 ♡23
fal/ideogram-v4-fast ♡23
t-tech/T-Search ♡23
GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking ♡23
mente-ai/uyu-2-28B ♡20
OpenMOSS-Team/MOSS-VL-Base-0708 ♡17
yijunwang2/krea2-outpaint ♡17
yijunwang2/krea2-reid ♡15
🦙 llama.cpp
└ Release: b9966 → b10068
└ 102 commits
- Added support for Hunyuan 3 (hy_v3) with MTP speculative decoding #25395 #25641
- Added support for Minimax2 Eagle3 speculative decoding 259ae1d
- Added support for BitNetForCausalLM GGUF conversion #25769
- Implemented GGML_OP_LIGHTNING_INDEXER for DeepSeek V3.2/V4 on CPU and CUDA #24231 #25545
- Added fused hyper-connection ops for DeepSeek V4 to reduce graph splits #25585 #25702
- Added CUDA Virtual Devices support and enabled CUDA graphs on Volta and Turing architectures #25228 #25749
- Added Flash Attention via oneDNN graph API for SYCL on Intel Battlemage #25222
- Optimized CUDA MoE gate/up activation quantization, improving prefill times on RTX 5090 and Blackwell #25441
- Added auto-download of DeepSeek-Flash and Eagle3 speculative decoding sidecars from Hugging Face #25811
- Server now supports CORS configuration options and accepts null sampling parameters to request defaults #25655 #25538
- Added KleidiAI SME2 f32 kernel and improved hardware-specific kernel dispatch #24414 #25478
- Fixed CUDA crash when querying memory on devices with no available memory #25157
- Fixed Tensor Parallel execution for Phi3, Bert, Plamo2/3, and ChatGLM #25536
- Fixed quantization crash on DeepSeek-V4 i32 routing tables #25787
🔗 All changes | Latest release
🎨 stable-diffusion.cpp
└ Release: master-775-b5d8120 → master-782-b290693
└ 7 commits
- Support for AnimateDiff SD 1.5 motion modules v2 and v3 #1784 with img2video capabilities via the --init-img parameter #1789
- Support for ADetailer #1785
- Support for PiD 1.5 #1790
- Configurable reference image processing for edit models #1780
- Fixed cross attention and output projection token protection for Anima LoRAs #1786
🔗 All changes | Latest release
🤗 Fresh models trending on HuggingFace:
thinkingmachines/Inkling ♡1060
OpenMOSS-Team/MOSS-VL-Realtime ♡76
ai-sage/GigaAM-Multilingual ♡56
nineninesix/diamond-1.0 ♡43
ai-sage/GigaChat3.1-Audio-10B-A1.8B ♡38
rzgar/Bernini-R-S2V ♡37
acvlab/ABot-World-0-5B-LF ♡29
fal/ideogram-v4-instant ♡27
InternScience/Agents-A1-4B ♡26
sensenova/SenseNova-U1-8B-MoT-Infographic-V3 ♡26
OpenMOSS-Team/MOSS-VL-Instruct-0708 ♡23
fal/ideogram-v4-fast ♡23
t-tech/T-Search ♡23
GnLOLot/MiniCPM5-1B-Claude-Opus-Fable5-V2-Thinking ♡23
mente-ai/uyu-2-28B ♡20
OpenMOSS-Team/MOSS-VL-Base-0708 ♡17
yijunwang2/krea2-outpaint ♡17
yijunwang2/krea2-reid ♡15
GitHub
model: add Hy3 (hy_v3) support with MTP speculative decoding by satindergrewal · Pull Request #25395 · ggml-org/llama.cpp
Overview
Adds support for Tencent's Hy3 (hy_v3 / HYV3ForCausalLM, 299B MoE, 80 layers + 1 MTP layer), including its multi-token-prediction head as a draft-mtp speculative target. Addresses ...
Adds support for Tencent's Hy3 (hy_v3 / HYV3ForCausalLM, 299B MoE, 80 layers + 1 MTP layer), including its multi-token-prediction head as a draft-mtp speculative target. Addresses ...
🔓 [HF Models] nvidia - Cosmos3-Edge
https://huggingface.co/nvidia/Cosmos3-Edge
🔓 [HF Models] nvidia - Cosmos3-Super-Text2Image-4Step
https://huggingface.co/nvidia/Cosmos3-Super-Text2Image-4Step
🔓 [HF Models] nvidia - Cosmos3-Super-Image2Video-4Step
https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step
🔓 [HF Models] nvidia - Cosmos3-Edge-Policy-DROID
https://huggingface.co/nvidia/Cosmos3-Edge-Policy-DROID
https://huggingface.co/nvidia/Cosmos3-Edge
🔓 [HF Models] nvidia - Cosmos3-Super-Text2Image-4Step
https://huggingface.co/nvidia/Cosmos3-Super-Text2Image-4Step
🔓 [HF Models] nvidia - Cosmos3-Super-Image2Video-4Step
https://huggingface.co/nvidia/Cosmos3-Super-Image2Video-4Step
🔓 [HF Models] nvidia - Cosmos3-Edge-Policy-DROID
https://huggingface.co/nvidia/Cosmos3-Edge-Policy-DROID
huggingface.co
nvidia/Cosmos3-Edge · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
📰 Google AI Blog - Run Ray on TPU, Part 1: The foundations
Ray 2.55 introduces official, first-class support for Google Cloud TPUs, enabling developers to run distributed Python workloads on Google's accelerators using the familiar Ray task-and-actor APIs. To handle the strict networking requirement of keeping multi-host TPU "slices" together over their Inter-Chip Interconnect (ICI), the KubeRay Operator on GKE automatically provisions and labels the underlying hardware layout. Ray Core utilizes these labels via its slice_placement_group() primitive to atomically reserve complete slices, allowing developers to deploy jobs through KubeRay, Ray Train, or Ray Serve simply by declaring a hardware topology (like "4x4") without writing custom placement code.
https://developers.googleblog.com/en/run-ray-on-tpu-part-1-the-foundations/
Ray 2.55 introduces official, first-class support for Google Cloud TPUs, enabling developers to run distributed Python workloads on Google's accelerators using the familiar Ray task-and-actor APIs. To handle the strict networking requirement of keeping multi-host TPU "slices" together over their Inter-Chip Interconnect (ICI), the KubeRay Operator on GKE automatically provisions and labels the underlying hardware layout. Ray Core utilizes these labels via its slice_placement_group() primitive to atomically reserve complete slices, allowing developers to deploy jobs through KubeRay, Ray Train, or Ray Serve simply by declaring a hardware topology (like "4x4") without writing custom placement code.
https://developers.googleblog.com/en/run-ray-on-tpu-part-1-the-foundations/
Googleblog
Google for Developers Blog - News about Web, Mobile, AI and Cloud
Ray 2.55 introduces official, first-class support for Google Cloud TPUs, enabling developers to run distributed Python workloads on Google's accelerators using the familiar Ray task-and-actor APIs. To handle the strict networking requirement of keeping multi…
📰 OpenAI - Safety and alignment in an era of long-horizon models
https://openai.com/index/safety-alignment-long-horizon-models
https://openai.com/index/safety-alignment-long-horizon-models
OpenAI
Safety and alignment in an era of long-horizon models
OpenAI shares lessons from deploying long-running AI models, highlighting new safety risks, observed failures, and improved safeguards through iterative deployment.
📰 HuggingFace - Grabette: an open system to record robot-manipulation data
https://huggingface.co/blog/grabette
https://huggingface.co/blog/grabette
huggingface.co
Grabette: an open system to record robot-manipulation data
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
📰 Google Model Cards - Gemini 3.6 Flash
https://deepmind.google/models/model-cards/gemini-3-6-flash/
📰 Google Model Cards - Gemini 3.5 Flash-Lite
https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/
https://deepmind.google/models/model-cards/gemini-3-6-flash/
📰 Google Model Cards - Gemini 3.5 Flash-Lite
https://deepmind.google/models/model-cards/gemini-3-5-flash-lite/
Google DeepMind
Gemini 3.6 Flash - Model Card
🆕 [HF Models] nvidia - NV-JEPA-DNA-HyenaDNA
https://huggingface.co/nvidia/NV-JEPA-DNA-HyenaDNA
🆕 [HF Models] nvidia - NV-JEPA-DNA-NTv3
https://huggingface.co/nvidia/NV-JEPA-DNA-NTv3
🆕 [HF Models] nvidia - NV-JEPA-DNA-DNABERT2
https://huggingface.co/nvidia/NV-JEPA-DNA-DNABERT2
https://huggingface.co/nvidia/NV-JEPA-DNA-HyenaDNA
🆕 [HF Models] nvidia - NV-JEPA-DNA-NTv3
https://huggingface.co/nvidia/NV-JEPA-DNA-NTv3
🆕 [HF Models] nvidia - NV-JEPA-DNA-DNABERT2
https://huggingface.co/nvidia/NV-JEPA-DNA-DNABERT2
huggingface.co
nvidia/NV-JEPA-DNA-HyenaDNA · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
📰 Google Gemma Blog - Scaling Agentic RL: High-Throughput Agentic Training with Tunix
Tunix is Google’s new JAX-native post-training library designed to eliminate TPU idling bottlenecks when training multi-turn, tool-using LLM reasoning agents. It maximizes hardware throughput by combining highly concurrent, asynchronous rollouts with a decoupled producer-consumer pipeline, ensuring the trainer is constantly fed even while agents wait on network I/O or environment steps. Additionally, Tunix provides plug-and-play abstractions and continuous macro-level profiling, allowing developers to easily integrate custom open-source environments and optimize complex distributed workflows without massive code rewrites.
https://developers.googleblog.com/en/scaling-agentic-rl-high-throughput-agentic-training-with-tunix/
Tunix is Google’s new JAX-native post-training library designed to eliminate TPU idling bottlenecks when training multi-turn, tool-using LLM reasoning agents. It maximizes hardware throughput by combining highly concurrent, asynchronous rollouts with a decoupled producer-consumer pipeline, ensuring the trainer is constantly fed even while agents wait on network I/O or environment steps. Additionally, Tunix provides plug-and-play abstractions and continuous macro-level profiling, allowing developers to easily integrate custom open-source environments and optimize complex distributed workflows without massive code rewrites.
https://developers.googleblog.com/en/scaling-agentic-rl-high-throughput-agentic-training-with-tunix/
Googleblog
Google for Developers Blog - News about Web, Mobile, AI and Cloud
Optimize agentic RL training with Tunix, Google’s JAX-native library. Eliminate TPU idling with async rollouts and easily plug in custom OSS environments.
📰 Claude Blog - How Anthropic secures its AI-native software development lifecycle
https://claude.com/blog/how-anthropic-secures-its-ai-native-software-development-lifecycle
📰 Claude Blog - How Datadog built a “universal machine tool” for Claude Code
https://claude.com/blog/how-datadog-built-a-universal-machine-tool-for-claude-code
📰 Claude Blog - Working at the frontier: How Rakuten builds agents overnight with Claude Fable 5
https://claude.com/blog/working-at-the-frontier-rakuten
https://claude.com/blog/how-anthropic-secures-its-ai-native-software-development-lifecycle
📰 Claude Blog - How Datadog built a “universal machine tool” for Claude Code
https://claude.com/blog/how-datadog-built-a-universal-machine-tool-for-claude-code
📰 Claude Blog - Working at the frontier: How Rakuten builds agents overnight with Claude Fable 5
https://claude.com/blog/working-at-the-frontier-rakuten
Claude
How Anthropic secures its AI-native software development lifecycle | Claude by Anthropic
Anthropic Deputy CISO Jason Clinton details how the Security Engineering team secures an AI-native SDLC where AI authors 80% of merged code.
📰 HuggingFace - The State of Simulation for Physical AI: An Overview
https://huggingface.co/blog/nvidia/state-of-simulation-for-physical-ai
https://huggingface.co/blog/nvidia/state-of-simulation-for-physical-ai
huggingface.co
The State of Simulation for Physical AI: An Overview
A Blog post by NVIDIA on Hugging Face
❤1
📰 NVIDIA - Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI
What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale.
https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/
📰 NVIDIA - NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI
Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context…
https://developer.nvidia.com/blog/inside-nvidia-vera-cpu-olympus-cores-built-for-maximum-single-threaded-performance-in-agentic-ai/
📰 NVIDIA - Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72
Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token…
https://developer.nvidia.com/blog/setting-a-world-record-for-moe-pre-training-on-nvidia-gb300-nvl72/
📰 NVIDIA - NVIDIA NVLink: The Scale-Up Network for AI Factories
The demand for AI continues to accelerate. Workloads are getting larger, models are becoming more complex, and there is mounting pressure to deploy AI compute…
https://developer.nvidia.com/blog/nvidia-nvlink-the-scale-up-network-for-ai-factories/
What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale.
https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai/
📰 NVIDIA - NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI
Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context…
https://developer.nvidia.com/blog/inside-nvidia-vera-cpu-olympus-cores-built-for-maximum-single-threaded-performance-in-agentic-ai/
📰 NVIDIA - Setting a World Record for MoE Pre-Training on NVIDIA GB300 NVL72
Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token…
https://developer.nvidia.com/blog/setting-a-world-record-for-moe-pre-training-on-nvidia-gb300-nvl72/
📰 NVIDIA - NVIDIA NVLink: The Scale-Up Network for AI Factories
The demand for AI continues to accelerate. Workloads are getting larger, models are becoming more complex, and there is mounting pressure to deploy AI compute…
https://developer.nvidia.com/blog/nvidia-nvlink-the-scale-up-network-for-ai-factories/
NVIDIA Technical Blog
Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI
What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale. These factories are now tasked with…
📰 Qwen Research - Qwen-Image-3.0: Rich Content, Authentic Details, Deep Knowledge
https://qwen.ai/blog?id=qwen-image-3.0
https://qwen.ai/blog?id=qwen-image-3.0
qwen.ai
Qwen Studio
Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.
🔄 [GitHub Releases] invoke-ai/InvokeAI - v6.13.7
https://github.com/invoke-ai/InvokeAI/releases/tag/v6.13.7
https://github.com/invoke-ai/InvokeAI/releases/tag/v6.13.7
GitHub
Release v6.13.7 · invoke-ai/InvokeAI
InvokeAI Version 6.13.7
⚠️ This is a security patch release
All recent versions of InvokeAI through 6.13.6 contain a security hole that potentially allows an attacker to recover the contents of the...
⚠️ This is a security patch release
All recent versions of InvokeAI through 6.13.6 contain a security hole that potentially allows an attacker to recover the contents of the...
❤1
📰 PyTorch - PyTorch Conference North America Schedule Is Live
PyTorch Conference North America will bring developers, researchers, and practitioners to San Jose on October 20–21 for sessions spanning training and inference, compiler innovations, responsible AI, applications, and the PyTorch...
https://pytorch.org/blog/pytorch-conference-north-america-schedule-is-live/
PyTorch Conference North America will bring developers, researchers, and practitioners to San Jose on October 20–21 for sessions spanning training and inference, compiler innovations, responsible AI, applications, and the PyTorch...
https://pytorch.org/blog/pytorch-conference-north-america-schedule-is-live/
🆕 [HF Models] microsoft - Mage-Flow-Edit-Base
https://huggingface.co/microsoft/Mage-Flow-Edit-Base
🆕 [HF Models] microsoft - Mage-Flow-Edit-Turbo
https://huggingface.co/microsoft/Mage-Flow-Edit-Turbo
🆕 [HF Models] microsoft - Mage-Flow-Edit
https://huggingface.co/microsoft/Mage-Flow-Edit
🆕 [HF Models] microsoft - Mage-Flow-Turbo
https://huggingface.co/microsoft/Mage-Flow-Turbo
🆕 [HF Models] microsoft - Mage-Flow
https://huggingface.co/microsoft/Mage-Flow
🆕 [HF Models] microsoft - Mage-Flow-Base
https://huggingface.co/microsoft/Mage-Flow-Base
https://huggingface.co/microsoft/Mage-Flow-Edit-Base
🆕 [HF Models] microsoft - Mage-Flow-Edit-Turbo
https://huggingface.co/microsoft/Mage-Flow-Edit-Turbo
🆕 [HF Models] microsoft - Mage-Flow-Edit
https://huggingface.co/microsoft/Mage-Flow-Edit
🆕 [HF Models] microsoft - Mage-Flow-Turbo
https://huggingface.co/microsoft/Mage-Flow-Turbo
🆕 [HF Models] microsoft - Mage-Flow
https://huggingface.co/microsoft/Mage-Flow
🆕 [HF Models] microsoft - Mage-Flow-Base
https://huggingface.co/microsoft/Mage-Flow-Base
huggingface.co
microsoft/Mage-Flow-Edit-Base · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.