APIUp to 28% cheaper than official pricesTry the API →
Hermes
// Intelligence

Today in AI

A focused feed of models, agents, research and open-source releases for people building with AI.

Updated
05
Google Drops Gemini Decision 1 to Score Choices 5x Cheaper Than Flash
newsApiLlmsGoogle AI Studio ·

Google Drops Gemini Decision 1 to Score Choices 5x Cheaper Than Flash

Google launched the Gemini Decisions API powered by new Gemini Decision 1 model. Scores up to 1,000 yes/no, choice, or score questions per request. Multimodal inputs: text, images, audio, video, and PDFs supported. Priced at $0.10 per 1M input tokens with zero output token charges.

06
LMSYS Doubles Muse Spark Open Speed to 1,109 Tokens per Second
newsInfraGpusLMSYS Org ·

LMSYS Doubles Muse Spark Open Speed to 1,109 Tokens per Second

SGLang megakernel takes Muse Spark Open from 517 to 1,109 tok/s on 8x GB300. Kernel launches per decode step drop from 1,450 to 97 by running each layer as one kernel. Weights prefetch into shared memory and L2; warp 3 is dedicated to warming the instruction cache.

10
Where Agent Capability Actually Comes From
deep_diveAgentsPost TrainingAkruti Acharya ·

Where Agent Capability Actually Comes From

Sharpening Tax in Post-Training evaluated 14 base/post-trained model pairs across three agentic benchmarks. Post-training generally improved first-attempt success ( pass@1 ), while base models often achieved greater coverage across repeated attempts ( pass@K ).

11
Comfy-Org Brings MiniMax H3-Image's 4K Local Generation to ComfyUI
modelImageOpen SourceAlphaSignal ·

Comfy-Org Brings MiniMax H3-Image's 4K Local Generation to ComfyUI

Comfy-Org published a repackaged, quantized MiniMax H3-Image build ready to drop into ComfyUI Repack ships int8 DiT, NVFP4 AWQ Qwen3-VL-32B text encoder, and fp16 VAE as single files Underlying MiniMax H3-Image is a 33.1B open-weight model generating 4096x4096 images Six sampling steps with no

12
Unsloth Shrinks Qwen3.8-27B to 6GB for Local Coding on Consumer Hardware
newsLlmsGpusUnsloth AI ·

Unsloth Shrinks Qwen3.8-27B to 6GB for Local Coding on Consumer Hardware

Unsloth released Dynamic V3.1 GGUFs for Qwen3.8-27B with top accuracy at every file size Leads rival quants by 5-10% on Terminal Bench, DeepSWE, and LiveCodeBench v6 Smallest 1-bit quant is 6.19 GB and runs on 12GB RAM machines Base model scores 73.0 Terminal Bench 2.1, 90.3 LiveCodeBench v6, 61.7

24
Edge0 Squeezes a 120B AI Agent Into one Consumer GPU
modelLlmsGpusAlphaSignal ·

Edge0 Squeezes a 120B AI Agent Into one Consumer GPU

Edge0 released 120B and 145B latent sparse MoE models trained natively at 4 bits. Built on Qwen3.8-Flash-Next; 512 experts per layer with 10 routed plus 1 shared active per token. 120B Q3_K hits 144.6 tok/s agent decode and 75.3 terminal-bench 2.1 on an RTX 5090 D.