🗓️ Weekly GitHub Activity


🦙 llama.cpp
└ Release: b10107 → b10229
└ 122 commits

- Added DSpark speculative decoding support on top of DFlash (#25173) and DeepSeek V4 MTP integration (#25784).
- Introduced support for Model Context Protocol (MCP) stdio transport in server (#26062).
- Added NextN/MTP speculative decoding support for GLM-5.2 (#25980).
- Added model support for MiniMax-M3 text and vision (#24908, #25113), GLM 5.2 / GLM-5.2-Vision (#25407, #26126), Nanbeige 4.2 (#25994), MiMo-V2.5 audio (#26190), Nemotron 3 Nano Omni (#22520), and Laguna-S-2.1 (#26233).
- CUDA backend added Q2_0 quantization support (#25707) and chunked SSD matmul for Mamba-2 prefill acceleration (#22675).
- SYCL backend added oneMKL GEMM flash attention for XMX acceleration (#25025) and RMS_NORM + MUL fusion (#26015).
- OpenCL backend now caches compiled binaries to disk (#26050).
- Added specialized chat parsers for MiniMax M3 (#26210) and Qwen3 (#26252).
- Fixed SYCL oneDNN flash attention scale memory corruption on long contexts (#25880).
- Fixed Metal memory leak when freeing models without GPU operations (#26082).

🔗 All changes | Latest release


🎨 stable-diffusion.cpp
└ Release: master-795-87a0177 → master-810-db99efd
└ 15 commits

- Added IP-Adapter Plus support with Resampler image projection (#1839)
- Exposed IP-Adapter parameters in server request schema and capabilities (#1824)
- Added support for Kroma-v0.1 LoRA models (#1842)
- Added linear multi-step sampling method (#1843)
- Allowed customizable alpha and beta parameters for the beta scheduler (#1834)

🔗 All changes | Latest release


🎵 audio.cpp
└ Release: release-0.4.2 → release-0.5
└ 74 commits

- Release 0.5 introducing new audio framework modules and model migrations (3178daf)
- Added AMD ROCm/HIP backend support for AMD GPU acceleration on Linux and Windows (#48, #153)
- Added live PCM audio streaming via CLI stdin and a live HTTP transcription endpoint (#118, #144)
- Added NVIDIA Parakeet-TDT 0.6B v3 ASR model (#111)
- Added Kroko Zipformer2 RNN-T ASR model with offline and stateful streaming support (#122)
- Added Fun-ASR-Nano offline ASR model (#155)
- Added Inflect Micro and Nano v2 TTS models (#125)
- Added BS-RoFormer source separation model (#114)
- Added Confucius4-TTS, DramaBox, and RVC models (#129)
- Added Qwen3 ASR streaming path (7ce769e)
- Accelerated Metal 1D transpose convolution execution for faster audio VAE decoding (#149)



🔗 All changes | Latest release

🤗 Fresh models trending on HuggingFace:

♡223 thinkingmachines/Inkling-Small 

A 276B parameter MoE multimodal LLM designed for conversational, coding, and agentic tasks across text, image, and audio inputs.

♡177 Audio8/Audio8-TTS-Preview-0.6b 

A 0.6B parameter multilingual TTS model optimized for zero-shot voice cloning and speech generation across 11 languages.

♡121 lodestones/Kroma 

A rank 256 style LoRA adapter and weight delta model for Krea 2

♡118 EschaLabs/Qwen3.6-35B-A3B-Escha-W2

2-bit quantized build of Qwen3.6-35B-A3B that outperforms FP8 quant

♡76 skt/A.X-K2

A 688B parameter MoE multilingual LLM featuring hybrid think/non-think reasoning for agentic tasks and long-context inference.

♡68 inference-optimization/Kimi-K3-0.40B 

A 0.40B parameter MoE LLM built as a tiny testing and development version of the Kimi-K3 architecture.

♡63 FermionResearch/Neutrino-8B

An 8B parameter ternary quantized chat LLM featuring sub-2-bit weights for highly efficient CPU and GPU inference.

♡62 KRAFTON/A.X-K2-Raon-Speech-21B-A3B —

A 21.2B parameter bilingual English/Korean speech LLM designed for STT, TTS, and multimodal turn-based chat

♡37 feyninc/FeyNobg

An image segmentation model designed for high-resolution background removal and alpha matting

♡36 sensenova/SenseNova-U1.5-8B-MoT-Preview

An 8B parameter multimodal image model designed for native 4K T2I generation and region-controllable image editing.

♡36 harrrshall/BarunLM-35M

A 35M parameter decoder-only base LLM designed for efficient local text generation and compact language model research.

♡36 Alissonerdx/CharacterSheet

A collection of image-editing LoRAs that converts character images into multi-view reference sheets for character design.

♡33 BAAI/AREX-Turbo

Finetune of Qwen-4B for for long-horizon agentic tasks and verification-guided research

♡32 lvladikov/SeedVR2-1.4B

A 1.44B parameter distilled one-step diffusion image upscaler designed for fast and resource-efficient super-resolution.

♡23 bench-labs/PixelModel-v5

A 40M parameter T2I diffusion transformer trained on 3 million captioned pairs for generating 256x256 images

♡21 OpenMOSS-Team/OmniVAE

An audio-video VAE and text-to-audio-video generation model designed for cross-modal aligned joint generation.

♡21 OrionLLM/GRM-3.2-Sky

A multimodal LLM optimized for long-horizon agentic workflows, coding, and mathematical reasoning

♡18 ProCreations/grug-3b

A 3B parameter reasoning LLM featuring caveman-style token-efficient reasoning.

♡18 openpangu/openPangu-2.0-Pro

A 505B-A18B parameter MoE LLM from Huawei supporting 512k context, trained on Huawei Ascend.

♡18 Baragi-AI/Munche-768

A 768-dimensional Korean fiction style embedding model based on Gemma 300M, designed for style similarity and authorship analysis.

♡16 Aratako/Irodori-TTS-v4-Small 

A Japanese Flow Diffusion Text-to-Speech model with zero-shot style-controlled voice cloning and text-based voice design.

🆕 [HF Models] internlm - Intern-S2-Preview-397B

https://huggingface.co/internlm/Intern-S2-Preview-397B


🔓 [HF Models] internlm - JanusCoder-14B

https://huggingface.co/internlm/JanusCoder-14B

🆕 [HF Models] MiniMaxAI - MiniMax-H3


https://huggingface.co/MiniMaxAI/MiniMax-H3

🔓 [HF Models] swiss-ai - wavtokenizer-large-unify-40token


https://huggingface.co/swiss-ai/wavtokenizer-large-unify-40token

📰 NVIDIA - NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage 
Storage is an active part of every agentic AI workflow. As agents retrieve enterprise knowledge, access persistent memory, reuse key-value (KV) cache data…

https://developer.nvidia.com/blog/nvidia-vera-storage-benchmarks-faster-encryption-compression-integrity-checking-and-recovery-for-ai-native-storage/

🆕 [HF Models] nvidia - NVIDIA-NemotronLabs-VoiceChat-11B


https://huggingface.co/nvidia/NVIDIA-NemotronLabs-VoiceChat-11B

📰 Qwen Research - Qwen3.8-Max: A New Bar for Coding and Cowork


https://qwen.ai/blog?id=qwen3.8

🆕 [HF Models] jdopensource - JoyAI-Video-Edit


https://huggingface.co/jdopensource/JoyAI-Video-Edit

📰 HuggingFace - Deploy local agents everywhere with LFM2.5-2.6B


https://huggingface.co/blog/LiquidAI/lfm2-5-2-6b

🆕 [HF Models] mistralai - Shieldstral-1.0-3B


https://huggingface.co/mistralai/Shieldstral-1.0-3B

🆕 [HF Models] LiquidAI - LFM2.5-2.6B-GGUF

https://huggingface.co/LiquidAI/LFM2.5-2.6B-GGUF


🆕 [HF Models] LiquidAI - LFM2.5-2.6B-Base

https://huggingface.co/LiquidAI/LFM2.5-2.6B-Base


🆕 [HF Models] LiquidAI - LFM2.5-2.6B

https://huggingface.co/LiquidAI/LFM2.5-2.6B

🆕 [HF Models] inclusionAI - Ling-3.0-flash


https://huggingface.co/inclusionAI/Ling-3.0-flash

📰 Mistral - Introducing Shieldstral.
Shieldstral introduces a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size.

https://mistral.ai/news/shieldstral/

🔓 [HF Models] nvidia - corrdiff-cosmo-era5


https://huggingface.co/nvidia/corrdiff-cosmo-era5

📰 Google DeepMind - The latest AI news we announced in July 2026


https://blog.google/innovation-and-ai/technology/ai/google-ai-updates-july-2026/

📰 Google AI Blog - A unified API for AI model routing
Google Cloud API Gateway now offers a model routing feature in Public Preview, allowing developers to dynamically route traffic to models like Gemini, Claude, or OpenAI OSS-GPT without hardcoding endpoints or managing open-source proxies. Developers can easily configure these routing rules directly within their OpenAPI 3.x specifications by mapping virtual model names to specific backend targets on a shared host. Once deployed, the Gateway acts as a serverless ingress layer that accepts standard OpenAI-compatible requests, automatically transcodes the payload to the native schema of the target model, and routes the traffic on the fly.

https://developers.googleblog.com/en/a-unified-api-for-ai-model-routing/

📰 LMSys - SpecForge v0.3.0: a Unified Disaggregated and Colocated Speculative Decoding Stack, and New Open SpecBundle Draft Models


https://lmsys.org/blog/2026-08-04-specforge-v0-3

📰 Anthropic - Mariano-Florentino (Tino) Cuéllar to join Anthropic as Chief Global Affairs Officer


https://www.anthropic.com/news/tino-cuellar

📰 OpenAI - New ways to learn and teach with ChatGPT Work and Codex
Explore new education plugins for ChatGPT Work and Codex that help K–12 teachers, college educators, and students learn, teach, research, and build.

https://openai.com/index/learn-teach-chatgpt-work-codex


📰 OpenAI - Apple is getting this wrong
OpenAI addresses Apple’s baseless lawsuit, corrects claims about its employees, and shares messages documenting what happened.

https://openai.com/index/apple-is-getting-this-wrong


📰 OpenAI - How we built a realtime system for responsive voice AI in six months
GPT-Live enables continuous voice interaction with AI, using a turnless speech model and low-latency architecture for faster, more natural conversations.

https://openai.com/index/continuous-voice-interaction-with-gpt-live

📰 NVIDIA - Beyond VLAs: How World Action Models Reshape Robot Manipulation
A central challenge in robotics is building policies that generalize beyond the demonstrations they’re trained on. A policy that succeeds in a training scene…

https://developer.nvidia.com/blog/beyond-vlas-how-world-action-models-reshape-robot-manipulation/