🔄 [GitHub Releases] huggingface/diffusers - Diffusers 0.39.0: New image and video pipelines, core library improvements, and more
https://github.com/huggingface/diffusers/releases/tag/v0.39.0
https://github.com/huggingface/diffusers/releases/tag/v0.39.0
GitHub
Release Diffusers 0.39.0: New image and video pipelines, core library improvements, and more · huggingface/diffusers
New Pipelines
Cosmos 3
Cosmos 3 is NVIDIA's unified world foundation model (WFM) for Physical AI — a single omni-model built on a Mixture-of-Transformers (MoT) architecture that combines world ...
Cosmos 3
Cosmos 3 is NVIDIA's unified world foundation model (WFM) for Physical AI — a single omni-model built on a Mixture-of-Transformers (MoT) architecture that combines world ...
📰 PyTorch - Understanding PyTorch’s Test Infrastructure
TL;DR PyTorch tests are often generated at import time, so CI failures may show device/dtype-specific names that differ from the source template. For local debugging, pytest -k and test/run_test.py are...
https://pytorch.org/blog/understanding-pytorchs-test-infrastructure/
TL;DR PyTorch tests are often generated at import time, so CI failures may show device/dtype-specific names that differ from the source template. For local debugging, pytest -k and test/run_test.py are...
https://pytorch.org/blog/understanding-pytorchs-test-infrastructure/
📰 Anthropic - More details on Fable 5’s cyber safeguards and our jailbreak framework
https://www.anthropic.com/news/fable-safeguards-jailbreak-framework
https://www.anthropic.com/news/fable-safeguards-jailbreak-framework
Anthropic
More details on Fable 5’s cyber safeguards and our jailbreak framework
What is and isn't blocked by our cyber classifiers, and a first draft of our jailbreak severity framework
🔄 [GitHub Releases] invoke-ai/InvokeAI - InvokeAI v6.13.5
https://github.com/invoke-ai/InvokeAI/releases/tag/v6.13.5
https://github.com/invoke-ai/InvokeAI/releases/tag/v6.13.5
GitHub
Release InvokeAI v6.13.5 · invoke-ai/InvokeAI
This is a maintenance release of InvokeAI focused on bug fixes and stability. Version 6.14.0 will be the next major feature release, featuring video generation, multiple GPU support, the Wan 2.2 im...
🗓️ Weekly GitHub Activity
🦙 llama.cpp
└ Release: b9828 → b9873
└ 45 commits
- Added model support and conversion for DeepSeek V4, featuring MoE gating adjustments, flash attention padding, and graph reuse #24162
- Added support for DFlash (v2) speculative decoding, including sliding window attention, draft model conversion, and speculative draft minimum probability parameters #22105, #25110, #25246
- Implemented a dedicated MiniCPM5 tool call parser and chat template support #24889
- Added llama_ftype_name to the public C API to retrieve model quantization type names #25134
- Reworked and optimized Hexagon Flash Attention, improving speed and accuracy #25085
- Added OpenCL support for Q1_0 quantization and precompiled binary kernel loading #25160, #23042
- Enabled CUDA topk-MoE fusion for models with 288 experts such as Step-3.7-Flash #25267
- Added CPU AVX2 optimizations for NVFP4 dot products #23961
- Integrated Model Context Protocol (MCP) server support and opt-in settings into the WebUI #25239
- Added SSE stream pinging to prevent connection drops on slow prefills in the server and WebUI #25241
- Added an --offline flag to llama download to verify local cache status without network requests #25091
- Fixed broken CPU concat implementation for quantized types #25247
- Fixed DFlash speculative decoding crashes by guarding unallocated K/V rotation buffers #25215
- Fixed infinite reasoning loops in StepFun by trimming messages sent to the parser #25238
🔗 All changes | Latest release
🎨 stable-diffusion.cpp
└ Release: master-721-8caa3f9 → master-749-b11c95a
└ 28 commits
- Added multi-device layer split, cross-device row split, and automatic tensor fitting to optimize load across devices #1734, #1735, #1736
- Added imatrix support #633
- Added Flux, Flux2, and Beta schedulers #1723, #1722, #811
- Added support for MiniT2I and SeFi-Image models #1683, #1707
- Added support for Qwen2D VAE, Qwen-Image/Wan VAE with diffusers naming, and Qwen image layered configurations #1714, #1713, #1119
- Enabled Wan/TAEHV video generation on the Metal backend #1731
- Added fallback handling when backend rejects flash attention #1732
- Improved third-party integration with installation support and pkg-config configuration #1632
🔗 All changes | Latest release
🤗 Fresh models trending on HuggingFace:
nationaldesignstudio/rampart ♡122
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 ♡48
kandinskylab/KVAE-Audio ♡47
Patil/Krea-2-depth-controlnet ♡36
kai-os/Grug-12B ♡34
Pageshift-Entertainment/pagestorm-research-preview-14b-full-book ♡28
openpangu/openPangu-2.0-Flash ♡28
Boogu/Boogu-Image-0.1-Edit-Turbo ♡27
prathoshap/vagdhenu ♡26
FINAL-Bench/Qwen3.5-35B-A3B-VKAE ♡26
RedHatAI/GLM-5.2-speculator.dspark ♡25
Aitrepreneur/IdeoKrea ♡25
wafer-bob/ASASR ♡23
RudySen/Krea2-realism-V2 ♡23
FINAL-Bench/metacog-adapter-Darwin-28B-Opus ♡23
FINAL-Bench/Darwin-36B-Opus-VKAE ♡23
malcolmrey/krea2 ♡22
FINAL-Bench/metacog-adapter-Ornith-1.0-9B ♡22
MaralGPT/MaralGPT-Mythos-9B-2606 ♡20
0xSero/GLM-5.2-504B-Nvidia ♡19
sensenova/SenseNova-U1-8B-MoT-Infographic-V2 ♡17
ai-sage/GFusion-10B-A1.8B ♡16
kai-os/Grug-35B-A3B ♡15
amalia-llm/AMALIA-9B-0626-SFT ♡15
TOTORONG/extGemma4-44B ♡14
🦙 llama.cpp
└ Release: b9828 → b9873
└ 45 commits
- Added model support and conversion for DeepSeek V4, featuring MoE gating adjustments, flash attention padding, and graph reuse #24162
- Added support for DFlash (v2) speculative decoding, including sliding window attention, draft model conversion, and speculative draft minimum probability parameters #22105, #25110, #25246
- Implemented a dedicated MiniCPM5 tool call parser and chat template support #24889
- Added llama_ftype_name to the public C API to retrieve model quantization type names #25134
- Reworked and optimized Hexagon Flash Attention, improving speed and accuracy #25085
- Added OpenCL support for Q1_0 quantization and precompiled binary kernel loading #25160, #23042
- Enabled CUDA topk-MoE fusion for models with 288 experts such as Step-3.7-Flash #25267
- Added CPU AVX2 optimizations for NVFP4 dot products #23961
- Integrated Model Context Protocol (MCP) server support and opt-in settings into the WebUI #25239
- Added SSE stream pinging to prevent connection drops on slow prefills in the server and WebUI #25241
- Added an --offline flag to llama download to verify local cache status without network requests #25091
- Fixed broken CPU concat implementation for quantized types #25247
- Fixed DFlash speculative decoding crashes by guarding unallocated K/V rotation buffers #25215
- Fixed infinite reasoning loops in StepFun by trimming messages sent to the parser #25238
🔗 All changes | Latest release
🎨 stable-diffusion.cpp
└ Release: master-721-8caa3f9 → master-749-b11c95a
└ 28 commits
- Added multi-device layer split, cross-device row split, and automatic tensor fitting to optimize load across devices #1734, #1735, #1736
- Added imatrix support #633
- Added Flux, Flux2, and Beta schedulers #1723, #1722, #811
- Added support for MiniT2I and SeFi-Image models #1683, #1707
- Added support for Qwen2D VAE, Qwen-Image/Wan VAE with diffusers naming, and Qwen image layered configurations #1714, #1713, #1119
- Enabled Wan/TAEHV video generation on the Metal backend #1731
- Added fallback handling when backend rejects flash attention #1732
- Improved third-party integration with installation support and pkg-config configuration #1632
🔗 All changes | Latest release
🤗 Fresh models trending on HuggingFace:
nationaldesignstudio/rampart ♡122
yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 ♡48
kandinskylab/KVAE-Audio ♡47
Patil/Krea-2-depth-controlnet ♡36
kai-os/Grug-12B ♡34
Pageshift-Entertainment/pagestorm-research-preview-14b-full-book ♡28
openpangu/openPangu-2.0-Flash ♡28
Boogu/Boogu-Image-0.1-Edit-Turbo ♡27
prathoshap/vagdhenu ♡26
FINAL-Bench/Qwen3.5-35B-A3B-VKAE ♡26
RedHatAI/GLM-5.2-speculator.dspark ♡25
Aitrepreneur/IdeoKrea ♡25
wafer-bob/ASASR ♡23
RudySen/Krea2-realism-V2 ♡23
FINAL-Bench/metacog-adapter-Darwin-28B-Opus ♡23
FINAL-Bench/Darwin-36B-Opus-VKAE ♡23
malcolmrey/krea2 ♡22
FINAL-Bench/metacog-adapter-Ornith-1.0-9B ♡22
MaralGPT/MaralGPT-Mythos-9B-2606 ♡20
0xSero/GLM-5.2-504B-Nvidia ♡19
sensenova/SenseNova-U1-8B-MoT-Infographic-V2 ♡17
ai-sage/GFusion-10B-A1.8B ♡16
kai-os/Grug-35B-A3B ♡15
amalia-llm/AMALIA-9B-0626-SFT ♡15
TOTORONG/extGemma4-44B ♡14
GitHub
DeepSeek V4 by am17an · Pull Request #24162 · ggml-org/llama.cpp
Overview
This PR adds support for the deepseek-v4 models. The most novel part of this architecture is the compressed attention. There are two types:
CSA (Compressed Sparse Attn) - it is a varia...
This PR adds support for the deepseek-v4 models. The most novel part of this architecture is the compressed attention. There are two types:
CSA (Compressed Sparse Attn) - it is a varia...
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_10-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_10-640
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_object-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_object-640
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_spatial-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_spatial-640
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_goal-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_goal-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_10-640
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_object-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_object-640
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_spatial-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_spatial-640
🆕 [HF Models] nvidia - gr00t17-lerobot-libero_goal-640
https://huggingface.co/nvidia/gr00t17-lerobot-libero_goal-640
huggingface.co
nvidia/gr00t17-lerobot-libero_10-640 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
🔄 [GitHub Releases] invoke-ai/InvokeAI - InvokeAI v6.13.6
https://github.com/invoke-ai/InvokeAI/releases/tag/v6.13.6
https://github.com/invoke-ai/InvokeAI/releases/tag/v6.13.6
GitHub
Release InvokeAI v6.13.6 · invoke-ai/InvokeAI
This is a patch release.
It corrects a bug introduced in v6.13.5 that caused Qwen Image and Qwen Image Edit models to crash with a dictionary key error when used with the default single-file Qwen 2...
It corrects a bug introduced in v6.13.5 that caused Qwen Image and Qwen Image Edit models to crash with a dictionary key error when used with the default single-file Qwen 2...
📰 HuggingFace - LeRobot v0.6.0: Imagine, Evaluate, Improve
https://huggingface.co/blog/lerobot-release-v060
https://huggingface.co/blog/lerobot-release-v060
📰 Google AI Blog - We terminated a TPU mid-training and it recovered in seconds: Introduction to elastic training with MaxText
Distributed AI training is notoriously fragile because losing a single machine typically crashes the entire multi-node job, forcing a time-consuming, full-workload infrastructure restart. To address this, Google’s JAX ecosystem utilizes elastic training via Pathways, which converts a hardware failure into a catchable Python exception so the running process can survive. When an unplanned failure occurs, the system automatically replaces only the broken worker, restores the last viable checkpoint from Cloud Storage, and resumes training in place—minimizing total downtime to under two minutes without ever restarting the main controller process.
https://developers.googleblog.com/en/we-terminated-a-tpu-mid-training-and-it-recovered-in-seconds-introduction-to-elastic-training-with-maxtext/
Distributed AI training is notoriously fragile because losing a single machine typically crashes the entire multi-node job, forcing a time-consuming, full-workload infrastructure restart. To address this, Google’s JAX ecosystem utilizes elastic training via Pathways, which converts a hardware failure into a catchable Python exception so the running process can survive. When an unplanned failure occurs, the system automatically replaces only the broken worker, restores the last viable checkpoint from Cloud Storage, and resumes training in place—minimizing total downtime to under two minutes without ever restarting the main controller process.
https://developers.googleblog.com/en/we-terminated-a-tpu-mid-training-and-it-recovered-in-seconds-introduction-to-elastic-training-with-maxtext/
Googleblog
Google for Developers Blog - News about Web, Mobile, AI and Cloud
Learn how elastic training on Cloud TPUs uses MaxText and Pathways to recover from multi-node hardware failures in under 2 minutes without restarting the job.
📰 LMSys - DSpark in SGLang: Speculative Decoding with Confidence-Driven, Variable-Length Verification
https://lmsys.org/blog/2026-07-06-dspark-sglang
https://lmsys.org/blog/2026-07-06-dspark-sglang
www.lmsys.org
DSpark in SGLang: Speculative Decoding with Confidence-Driven, Variable-Length Verification
Speculative decoding trades extra compute for fewer decode steps, and the trade
sours as load grows: at batch size B with K speculative tokens the target
verifies B K tokens every step, and past a poi...
sours as load grows: at batch size B with K speculative tokens the target
verifies B K tokens every step, and past a poi...
📰 Anthropic Research - A global workspace in language models
https://www.anthropic.com/research/global-workspace
https://www.anthropic.com/research/global-workspace
Anthropic
A global workspace in language models
Interpretability research on Claude's internal thoughts.