📰 NVIDIA - Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading
Large language model (LLM) training workloads increasingly run into GPU memory limits before compute is fully used. Model weights, gradients, optimizer states…
https://developer.nvidia.com/blog/reducing-high-bandwidth-memory-bottlenecks-in-jax-based-llm-training-with-host-offloading/
📰 NVIDIA - AI Model Co-Design: Hardware-Friendly LLM Design
AI performance comes down to three dimensions: Deployments must balance all three: High accuracy is wasted if responses are slow, and raw throughput means…
https://developer.nvidia.com/blog/ai-model-co-design-hardware-friendly-llm-design/
📰 NVIDIA - Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit
Biomolecular structure prediction and co-folding with models like OpenFold3 are now mainstream, large-scale workloads powering drug discovery and protein design.
https://developer.nvidia.com/blog/accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-agent-toolkit/
Large language model (LLM) training workloads increasingly run into GPU memory limits before compute is fully used. Model weights, gradients, optimizer states…
https://developer.nvidia.com/blog/reducing-high-bandwidth-memory-bottlenecks-in-jax-based-llm-training-with-host-offloading/
📰 NVIDIA - AI Model Co-Design: Hardware-Friendly LLM Design
AI performance comes down to three dimensions: Deployments must balance all three: High accuracy is wasted if responses are slow, and raw throughput means…
https://developer.nvidia.com/blog/ai-model-co-design-hardware-friendly-llm-design/
📰 NVIDIA - Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit
Biomolecular structure prediction and co-folding with models like OpenFold3 are now mainstream, large-scale workloads powering drug discovery and protein design.
https://developer.nvidia.com/blog/accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-agent-toolkit/
NVIDIA Technical Blog
Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading
Large language model (LLM) training workloads increasingly run into GPU memory limits before compute is fully used. Model weights, gradients, optimizer states, communication buffers…
🔄 [GitHub Releases] sgl-project/sglang - v0.5.15
https://github.com/sgl-project/sglang/releases/tag/v0.5.15
https://github.com/sgl-project/sglang/releases/tag/v0.5.15
GitHub
Release v0.5.15 · sgl-project/sglang
Highlights
GLM-5.2 NVFP4, tuned for production: We took time this cycle to tune GLM-5.2 NVFP4 on Blackwell for optimized production serving. It now runs at 500+ tok/s/user on 8x B300, 450 on 4x GB3...
GLM-5.2 NVFP4, tuned for production: We took time this cycle to tune GLM-5.2 NVFP4 on Blackwell for optimized production serving. It now runs at 500+ tok/s/user on 8x B300, 450 on 4x GB3...
🆕 [HF Models] inclusionAI - SingGuard-NSFA-9B
https://huggingface.co/inclusionAI/SingGuard-NSFA-9B
🆕 [HF Models] inclusionAI - SingGuard-NSFA-4B
https://huggingface.co/inclusionAI/SingGuard-NSFA-4B
🆕 [HF Models] inclusionAI - SingGuard-NSFA-2B
https://huggingface.co/inclusionAI/SingGuard-NSFA-2B
🆕 [HF Models] inclusionAI - SingGuard-NSFA-0.8B
https://huggingface.co/inclusionAI/SingGuard-NSFA-0.8B
https://huggingface.co/inclusionAI/SingGuard-NSFA-9B
🆕 [HF Models] inclusionAI - SingGuard-NSFA-4B
https://huggingface.co/inclusionAI/SingGuard-NSFA-4B
🆕 [HF Models] inclusionAI - SingGuard-NSFA-2B
https://huggingface.co/inclusionAI/SingGuard-NSFA-2B
🆕 [HF Models] inclusionAI - SingGuard-NSFA-0.8B
https://huggingface.co/inclusionAI/SingGuard-NSFA-0.8B
huggingface.co
inclusionAI/SingGuard-NSFA-9B · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
🔄 [GitHub Releases] vllm-project/vllm - v0.25.0
https://github.com/vllm-project/vllm/releases/tag/v0.25.0
https://github.com/vllm-project/vllm/releases/tag/v0.25.0
GitHub
Release v0.25.0 · vllm-project/vllm
vLLM v0.25.0 Release Notes
Highlights
This release features 558 commits from 232 contributors (64 new)!
Model Runner V2 is now the default for all dense models (#44443). Building on quantized-mode...
Highlights
This release features 558 commits from 232 contributors (64 new)!
Model Runner V2 is now the default for all dense models (#44443). Building on quantized-mode...
🗓️ Weekly GitHub Activity
🦙 llama.cpp
└ Release: b9873 → b9966
└ 93 commits
- Added initial support for the ET backend targeting ET-Soc-1 hardware #24179
- Introduced Q2_0 quantization format with CPU backend support for Ternary Bonsai models #24448
- Added multimodal support for DeepSeek-OCR v1 multi-tile dynamic resolution #24717
- Refactored llama-cli into an HTTP-based implementation interacting with the server #24948
- Fused MMVQ post-scale for NVFP4 on CUDA to accelerate FP8 and NVFP4 models #24481
- Optimized OpenCL Flash Attention decoding performance #25366
- Fixed a decode bottleneck in tensor-split mode by compiling regex patterns statically #24710
- Enabled unsafe math optimizations for AMD/HIP builds to match CUDA performance #24668
- Fixed a security vulnerability involving out-of-bounds reads in the UGM tokenizer #18750
- Fixed a crash occurring when using tensor parallelism with CPU-offloaded MoE experts #25028
🔗 All changes | Latest release
🎨 stable-diffusion.cpp
└ Release: master-749-b11c95a → master-775-b5d8120
└ 26 commits
- Added support for Krea2OstrisEdit #1775 and lingbot video #1770
- Support hot-reloading ControlNet to swap models without rebuilding the context #1768
- Added DPM++ 2M SDE and DPM++ 2M SDE Brownian tree samplers #1742 and #1743
- Support loading safetensors index files #1769
- Use denoise strength as the starting noise level #1738
- Moved circular padding from context to per-generation parameters #1748
- Improve generation speed by driving layer splitting from graph-cut segments #1762
- Fixed SDXL ControlNet integration issues regarding diffusers naming and graph size #1752
- Fixed UNet block paths splitting across layers #1741
🔗 All changes | Latest release
🤗 Fresh models trending on HuggingFace:
bottlecapai/ThinkingCap-Qwen3.6-27B ♡235
conradlocke/krea2-identity-edit ♡185
Alissonerdx/LTX-Best-Face-ID ♡99
SupraLabs/Supra-Router-51M ♡98
Patil/Krea-2-depth-controlnet ♡91
robbyant/lingbot-video-moe-30b-a3b ♡84
migtissera/Tess-4-27B ♡84
mgwr/M87 ♡68
ostris/krea2_turbo_style_reference ♡62
ai-sage/GigaChat3.5-432B-A28B ♡60
robbyant/lingbot-world-v2-14b-causal-fast ♡60
empero-ai/Qwythos-9B-v2 ♡43
wikeeyang/Krea2-Turbo-HD-V1 ♡28
MirilAI/Miril-Drone-2B-1 ♡27
robbyant/lingbot-vla-v2-6b ♡27
robbyant/lingbot-video-dense-1.3b ♡23
Ateron/Gemma-4-Novelist-Eclipse-31B ♡23
rzgar/Bernini-R-S2V ♡21
SOLRICKS/ltx-2.3-product-ad-style ♡21
ostris/Krea2OstrisEdit ♡20
ai-sage/GigaChat3.5-432B-A28B-base ♡18
robbyant/lingbot-vision-vit-large ♡18
epfl-neuroai/NEvo ♡18
OrionLLM/GRM-2.6-Plus-0628 ♡17
FrontiersMind/Lumma-0.6B-Base ♡13
sais-org/Polaris_Pro ♡13
🦙 llama.cpp
└ Release: b9873 → b9966
└ 93 commits
- Added initial support for the ET backend targeting ET-Soc-1 hardware #24179
- Introduced Q2_0 quantization format with CPU backend support for Ternary Bonsai models #24448
- Added multimodal support for DeepSeek-OCR v1 multi-tile dynamic resolution #24717
- Refactored llama-cli into an HTTP-based implementation interacting with the server #24948
- Fused MMVQ post-scale for NVFP4 on CUDA to accelerate FP8 and NVFP4 models #24481
- Optimized OpenCL Flash Attention decoding performance #25366
- Fixed a decode bottleneck in tensor-split mode by compiling regex patterns statically #24710
- Enabled unsafe math optimizations for AMD/HIP builds to match CUDA performance #24668
- Fixed a security vulnerability involving out-of-bounds reads in the UGM tokenizer #18750
- Fixed a crash occurring when using tensor parallelism with CPU-offloaded MoE experts #25028
🔗 All changes | Latest release
🎨 stable-diffusion.cpp
└ Release: master-749-b11c95a → master-775-b5d8120
└ 26 commits
- Added support for Krea2OstrisEdit #1775 and lingbot video #1770
- Support hot-reloading ControlNet to swap models without rebuilding the context #1768
- Added DPM++ 2M SDE and DPM++ 2M SDE Brownian tree samplers #1742 and #1743
- Support loading safetensors index files #1769
- Use denoise strength as the starting noise level #1738
- Moved circular padding from context to per-generation parameters #1748
- Improve generation speed by driving layer splitting from graph-cut segments #1762
- Fixed SDXL ControlNet integration issues regarding diffusers naming and graph size #1752
- Fixed UNet block paths splitting across layers #1741
🔗 All changes | Latest release
🤗 Fresh models trending on HuggingFace:
bottlecapai/ThinkingCap-Qwen3.6-27B ♡235
conradlocke/krea2-identity-edit ♡185
Alissonerdx/LTX-Best-Face-ID ♡99
SupraLabs/Supra-Router-51M ♡98
Patil/Krea-2-depth-controlnet ♡91
robbyant/lingbot-video-moe-30b-a3b ♡84
migtissera/Tess-4-27B ♡84
mgwr/M87 ♡68
ostris/krea2_turbo_style_reference ♡62
ai-sage/GigaChat3.5-432B-A28B ♡60
robbyant/lingbot-world-v2-14b-causal-fast ♡60
empero-ai/Qwythos-9B-v2 ♡43
wikeeyang/Krea2-Turbo-HD-V1 ♡28
MirilAI/Miril-Drone-2B-1 ♡27
robbyant/lingbot-vla-v2-6b ♡27
robbyant/lingbot-video-dense-1.3b ♡23
Ateron/Gemma-4-Novelist-Eclipse-31B ♡23
rzgar/Bernini-R-S2V ♡21
SOLRICKS/ltx-2.3-product-ad-style ♡21
ostris/Krea2OstrisEdit ♡20
ai-sage/GigaChat3.5-432B-A28B-base ♡18
robbyant/lingbot-vision-vit-large ♡18
epfl-neuroai/NEvo ♡18
OrionLLM/GRM-2.6-Plus-0628 ♡17
FrontiersMind/Lumma-0.6B-Base ♡13
sais-org/Polaris_Pro ♡13
GitHub
Initial ET backend by marty1885 · Pull Request #24179 · ggml-org/llama.cpp
Overview
This PR is developed by AINekko and by members of AIFoundry (AINekko's OSS community) and adds the ET backend that supports the ET-SOC-1 processor. ET-SOC-1 was originally created ...
This PR is developed by AINekko and by members of AIFoundry (AINekko's OSS community) and adds the ET backend that supports the ET-SOC-1 processor. ET-SOC-1 was originally created ...
📰 Claude Blog - Working at the frontier: How Hebbia builds AI for financial diligence that can't miss a detail
https://claude.com/blog/working-at-the-frontier-how-hebbia-builds-ai-for-financial-diligence-that-cant-miss-a-detail
https://claude.com/blog/working-at-the-frontier-how-hebbia-builds-ai-for-financial-diligence-that-cant-miss-a-detail
Claude
Working at the frontier: How Hebbia builds AI for financial diligence that can't miss a detail | Claude by Anthropic
How Anthropic's Claude Fable 5 beat Hebbia's finance-specific model evaluations, achieving their biggest accuracy gain yet.
📰 Anthropic Research - Claude’s values across models and languages
https://www.anthropic.com/research/claude-values-models-languages
📰 Anthropic Research - Claude plays robotics
https://www.anthropic.com/research/claude-plays-robotics
https://www.anthropic.com/research/claude-values-models-languages
📰 Anthropic Research - Claude plays robotics
https://www.anthropic.com/research/claude-plays-robotics
Anthropic
Claude’s values across models and languages
We analyzed 300,000 real conversations to measure the values Claude expresses across models and languages, compressed into four interpretable axes.
📰 NVIDIA - NVIDIA Ising Decoding Cuts Color Code Logical Error Rates by Over 300X
Useful quantum computers will require fault tolerant logical operations. Researchers are actively exploring many different quantum error correction (QEC) codes…
https://developer.nvidia.com/blog/nvidia-ising-decoding-cuts-color-code-logical-error-rates-by-over-300x/
Useful quantum computers will require fault tolerant logical operations. Researchers are actively exploring many different quantum error correction (QEC) codes…
https://developer.nvidia.com/blog/nvidia-ising-decoding-cuts-color-code-logical-error-rates-by-over-300x/
NVIDIA Technical Blog
NVIDIA Ising Decoding Cuts Color Code Logical Error Rates by Over 300X
Useful quantum computers will require fault tolerant logical operations. Researchers are actively exploring many different quantum error correction (QEC) codes to enable this…
🆕 [HF Models] inclusionAI - SingGuard-NSFA-9B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-9B-GGUF
🆕 [HF Models] inclusionAI - SingGuard-NSFA-4B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-4B-GGUF
🆕 [HF Models] inclusionAI - SingGuard-NSFA-2B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-2B-GGUF
🆕 [HF Models] inclusionAI - SingGuard-NSFA-0.8B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-0.8B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-9B-GGUF
🆕 [HF Models] inclusionAI - SingGuard-NSFA-4B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-4B-GGUF
🆕 [HF Models] inclusionAI - SingGuard-NSFA-2B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-2B-GGUF
🆕 [HF Models] inclusionAI - SingGuard-NSFA-0.8B-GGUF
https://huggingface.co/inclusionAI/SingGuard-NSFA-0.8B-GGUF
huggingface.co
inclusionAI/SingGuard-NSFA-9B-GGUF · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
🔄 [GitHub Releases] sgl-project/sglang - v0.5.15.post1
https://github.com/sgl-project/sglang/releases/tag/v0.5.15.post1
https://github.com/sgl-project/sglang/releases/tag/v0.5.15.post1
GitHub
Release v0.5.15.post1 · sgl-project/sglang
v0.5.15.post1 includes a few patches, mostly for GLM 5.2
#30454 #30627: Fix DSA model launching on non Cuda/HIP devices
#30858: Fix flashinfer dependency on Cuda 12 images
#31001: Fix NaN outputs ...
#30454 #30627: Fix DSA model launching on non Cuda/HIP devices
#30858: Fix flashinfer dependency on Cuda 12 images
#31001: Fix NaN outputs ...
🔄 [GitHub Releases] vllm-project/vllm - v0.25.1
https://github.com/vllm-project/vllm/releases/tag/v0.25.1
https://github.com/vllm-project/vllm/releases/tag/v0.25.1
GitHub
Release v0.25.1 · vllm-project/vllm
vLLM v0.25.1
Highlights
This release features 2 commits from 2 contributors (1 new)!
v0.25.1 is a patch release containing two targeted bug fixes on top of v0.25.0.
Bug Fixes
Avoid blocking model ...
Highlights
This release features 2 commits from 2 contributors (1 new)!
v0.25.1 is a patch release containing two targeted bug fixes on top of v0.25.0.
Bug Fixes
Avoid blocking model ...
🆕 [HF Models] nvidia - Nemotron-3-Embed-8B-BF16
https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16
🆕 [HF Models] nvidia - Nemotron-3-Embed-1B-BF16
https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16
https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16
🆕 [HF Models] nvidia - Nemotron-3-Embed-1B-BF16
https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16
huggingface.co
nvidia/Nemotron-3-Embed-8B-BF16 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
🔄 [GitHub Releases] turboderp-org/exllamav3 - 1.0.0
https://github.com/turboderp-org/exllamav3/releases/tag/v1.0.0
https://github.com/turboderp-org/exllamav3/releases/tag/v1.0.0
GitHub
Release 1.0.0 · turboderp-org/exllamav3
-> Small writeup with charts.
Remove flash-attention-2 and xformers dependencies
New attention kernel with online cache quantization, dual input for SWA layers and attention sinks
New conv1d ke...
Remove flash-attention-2 and xformers dependencies
New attention kernel with online cache quantization, dual input for SWA layers and attention sinks
New conv1d ke...
📰 Google DeepMind - Reconstructing Pelé’s “lost” goal
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/reconstructing-peles-lost-goal/
https://blog.google/innovation-and-ai/models-and-research/google-deepmind/reconstructing-peles-lost-goal/
Google
Reconstructing Pelé’s “lost” goal
See how Google DeepMind AI technology reconstructed Pelé’s legendary 1959 lost goal at Rua Javari in our new mini-documentary.