📰 PyTorch - PyTorch Conference North America Announces 2026 Keynotes
PyTorch Conference North America will be held in San Jose, California, on October 20–21, 2026. The 2026 keynote speakers are: Mark Collier, Executive Director, PyTorch Foundation Mazin Gilbert, Executive Director,...

https://pytorch.org/blog/pytorch-conference-north-america-announces-2026-keynotes/

🔄 [GitHub Releases] turboderp-org/exllamav3 - 1.4.0


https://github.com/turboderp-org/exllamav3/releases/tag/v1.4.0

🆕 [HF Models] ATH-MaaS - Lumen


https://huggingface.co/ATH-MaaS/Lumen


🆕 [HF Models] ATH-MaaS - Lumen_wmt26


https://huggingface.co/ATH-MaaS/Lumen_wmt26

🆕 [HF Models] Wan-AI - Wan2.2-Animate-2-14B-Distilled-Diffusers

https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B-Distilled-Diffusers


🆕 [HF Models] Wan-AI - Wan2.2-Animate-2-14B-Diffusers

https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B-Diffusers


🆕 [HF Models] Wan-AI - Wan2.2-Animate-2-14B

https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B

📰 Google DeepMind - See what 5 builders are making with Gemini Omni


https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni-builders/

📰 LMSys - HPC-Ops × SGLang: High-Performance Attention, Router GEMM, and MoE Kernels from Tencent Hunyuan


https://lmsys.org/blog/2026-08-07-hpc-ops-sglang

📰 HuggingFace - TutorMoments: Do AI tutors know when to help and when to hold back?


https://huggingface.co/blog/allenai/tutormoments

📰 Anthropic - Improving Fable 5's biology safeguards


https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards

📰 OpenAI - Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users
ChatGPT introduces improved GPT-5.6 Sol with better accuracy and consistency, plus expanded access for free users and unlimited everyday chats with GPT-5.6 Luna.

https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt


📰 OpenAI - Working with the American Psychological Association on youth mental health and AI
OpenAI and the American Psychological Association advance evidence-based guidance, resources, and safeguards for responsible AI use and youth mental health.

https://openai.com/index/openai-and-apa-partner-to-advance-responsible-ai


📰 OpenAI - From asking to doing: How the world is putting ChatGPT to work
New OpenAI Signals data shows how people use ChatGPT worldwide, with country-level insights on adoption, usage trends, and evolving behavior.

https://openai.com/index/how-the-world-is-putting-chatgpt-to-work

🔄 [GitHub Releases] sgl-project/sglang - v0.5.17


https://github.com/sgl-project/sglang/releases/tag/v0.5.17

🆕 [GitHub Repos] orgs/NVlabs - RoboVoLo


https://github.com/NVlabs/RoboVoLo

📰 xAI - Imagine Image 2.0


https://x.ai/news/grok-imagine-image-2

🔄 [GitHub Releases] turboderp-org/exllamav3 - 1.4.1


https://github.com/turboderp-org/exllamav3/releases/tag/v1.4.1

📰 Claude Blog - Auto mode is now the default in Claude Code for Pro, Max, and Team plans

https://claude.com/blog/auto-mode-default-in-claude-code


📰 Claude Blog - Running auto mode in production

https://claude.com/blog/auto-mode-in-production


📰 Claude Blog - Millennium and Anthropic are building a digital risk analyst with Claude

https://claude.com/blog/millennium-and-anthropic-are-building-a-digital-risk-analyst-with-claude


📰 Claude Blog - Run Claude Code sessions on your own compute

https://claude.com/blog/run-claude-code-sessions-on-your-own-compute


📰 Claude Blog - Inference hooks: inline data loss prevention for Claude Enterprise

https://claude.com/blog/claude-enterprise-inference-hooks


📰 Claude Blog - A guide to cost visibility and control in Claude

https://claude.com/blog/a-guide-to-cost-visibility-and-control-in-claude

🗓️ Weekly GitHub Activity


🦙 llama.cpp
└ Release: b10233 → b10331
└ 98 commits

- Added Multi-Token Prediction (MTP) support for DeepSeek V3.2 #26457 and Qwen3-Next #25589
- Added support for Qwen3-TTS #26254 and DeepSeek OCR multi-row batching #26154
- Introduced Docker-based tool isolation in the server #26507
- Added an LRU scheduler to the server router to prevent evicting busy models #26572
- Upgraded the WebUI with filesystem @mentions #26715, slash commands #26716, a working directory picker #26518, and rich contenteditable input #26717
- Implemented the DeepSeek V4 Lightning Indexer on Metal #25893 and SYCL #26568 to boost prompt processing performance
- Fused CUDA kernels for rms_norm + mul + rope to improve inference latency #26767
- Enabled dynamic allocation for split graph inputs to prevent crashes when running wide MoE models #22789
- Extended SYCL oneDNN SDPA to non-FP16 KV caches (Q4_0-Q8_0, FP32) #25874
- Added a --delete-splits option to gguf-split to free up disk space during file merges #26538
- Added an upcoming deprecation notice for changing the default server port from 8080 to 9931 #26508
- Fixed CUDA data-races when reusing shared memory in block_reduce #26385
- Fixed Metal NORM/RMS_NORM calculation errors on row lengths leaving a partial simdgroup #26708
- Fixed gguf-py reader to guard against OOM and integer overflow on crafted files #25401

🔗 All changes | Latest release


🎨 stable-diffusion.cpp
└ Release: master-810-db99efd → master-813-bfbef5b
└ 3 commits

- Added support for the minimax-h3 model #1854
- Added a trained Minimax VAE Latent2rgb projection #1856

🔗 All changes | Latest release


🎵 audio.cpp
└ Release: release-0.5 → release-0.5.1
└ 12 commits

- Added a --list-devices CLI flag and list_backend_devices() API to enumerate available backend devices with their hardware names #171
- Added support for Irodori-TTS v4 and aligned its tokenizer #184
- Migrated multiple models to spec v1, including HTDemucs, Mel-Band RoFormer, HeartMuLa, Hviske ASR, SeedVC, Sortformer, and Higgs Audio STT #177 #179
- Enabled Metal ConvTranspose fast path with MioCodec layout-safe multiplies cdd5196

🔗 All changes | Latest release

🤗 Fresh models trending on HuggingFace:

♡255 deepgrove/maple-preview 

20B-A1B ternary-weight Mixture-of-Experts LLM optimized for high-speed on-device reasoning and math.

♡197 lightx2v/Minimax-h3-Turbo

Image-to-video and text-to-video model based on MiniMax-H3 for fast video generation.

♡168 PinkCherry_MiniMax-H3 | PinkFluffyBunny-MiniMax-H3

Text-to-video model based on MiniMax-H3 specialized in organic motion and detail generation.

♡139 Kijai/MiniMax-H3-experimental

Experimental 4-bit quantized version of the MiniMax-H3 video model with int8-convrot activation support for ComfyUI.

♡97 SyzygyResearch/Mach-1-Additive-35B

35B ternary 1.7-bit LLM based on Qwen designed for highly efficient, high-accuracy reasoning.

♡94 Kijai/MiniMax-H3-TAE

Lightweight VAE model designed for quick video preview generation in MiniMax-H3 workflows.

♡93 endless-frontier/BigBang-v1

Qwen3.6-35B-A3B fine-tune for synthetic frontier tasks for advanced scientific research, coding, and tool-use, significantly outperforming base model on benchmarks. [paper]

♡81 Akahsizrr/fuse-1-Lite 

5.72B MoE LLM fusing LiquidAI's host model with Qwen3.6 coding experts for lightweight code generation.

♡75 badtheorylabs/BTL-4 | BTL-4-Compact

35B agentic reasoning LLM fine-tuned for tool use, programming, and long-horizon software engineering tasks.

♡49 harrrshall/BarunLM-35M

35M parameter decoder-only base SLM featuring hybrid local-global attention for ultra-efficient text generation.

♡44 TenStrip/10Eros-Max

Video diffusion model based on MiniMax-H3 that blends weight patterns from LTX, Wan, and Krea models to improve aesthetic detail.

♡42 sand-ai/MAGI-2-preview

114B MoE audio-video generator activating 6B parameters per token to produce synchronized 10-second videos with sound.

♡40 thedeoxen/Krea-2-pose-controlnet

OpenPose ControlNet LoRA for Krea2 Turbo that controls character posing based on input skeleton maps.

♡39 ReadyArt/gemma-4-31B-it-scotoma-2 

31B LLM fine-tuned from Gemma 4 to reduce stylistic ai tics and cautious behavioral constraints during generation.

♡39 DeepBeepMeep/MiniMax-H3

Single-file video and image generation models compatible with the WanGP interface for low-VRAM deployment.

♡35 zerofata/G4-MeroMero-v2-31B

31B Creative narrative & RP tune of Gemma 4.

♡32 SupraLabs/Supra2-100M-Instruct | Supra2-100M-Base 

100M parameter instruct-tuned SLM based on the Qwen3 architecture with a 2K token context window.

♡29 KBlueLeaf/TIPOv2-1B-A200M

1B MoE sparse prompt expansion model designed to optimize short user prompts for T2I generators.

♡27 ASLP-lab/MeanVC2

18M parameter zero-shot voice conversion model designed for ultra-low-latency streaming audio.

♡26 quimmedes/Deepwen-3.6

35B MoE LLM fine-tuned with DeepSeek-style reasoning and specialized skills for 2D and 3D game development and Blender pipelines.

♡19 BananaMind/BananaMind-2-Pro-Preview

139M parameter base SLM preview checkpoint featuring a digit-aware BPE tokenizer and a 3K token context.

♡19 0xSero/deepseek-v4-flash-0731-spark 

REAP-pruned, quantized MoE LLM optimized for ultra-long context inference on a single DGX Spark.

🔓 [HF Models] microsoft - MuseVLA


https://huggingface.co/microsoft/MuseVLA

🆕 [HF Models] dots-studio - dots.tts-mf-2steps-interleave-tn

https://huggingface.co/dots-studio/dots.tts-mf-2steps-interleave-tn


🆕 [HF Models] dots-studio - dots.tts-mf-2steps

https://huggingface.co/dots-studio/dots.tts-mf-2steps


🆕 [HF Models] dots-studio - dots.tts-mf-1step

https://huggingface.co/dots-studio/dots.tts-mf-1step


🆕 [HF Models] dots-studio - dots.tts-mf-2steps-stts


https://huggingface.co/dots-studio/dots.tts-mf-2steps-stts

📰 HuggingFace - Making Knowledge Distillation Cheap Enough to Run at Scale


https://huggingface.co/blog/MultiverseComputingCAI/efficient-knowledge-distillation

🆕 [HF Models] Motif-Technologies - Motif-3-Base

https://huggingface.co/Motif-Technologies/Motif-3-Base


🆕 [HF Models] Motif-Technologies - Motif-3

https://huggingface.co/Motif-Technologies/Motif-3

🆕 [HF Models] meta-models - Muse-Glimmer-30B

https://huggingface.co/meta-models/Muse-Glimmer-30B | GGUF