Quantized LLMs recover 97% accuracy; IBM adds reasoning to Granite

Quantization-Aware Distillation lets Q4_0 GGUF models retain about 97% of BF16 accuracy at 4-bit speed and memory. IBM released open Granite 4.2 models (3B, 8B, 30B) with chain-of-thought reasoning, native tool calling and a 512K-token context window.

DEV Community

Open source

Quantized LLMs recover 97% accuracy; IBM adds reasoning to Granite

QAD Q4_0 GGUF models recover ~97% of BF16 accuracy with no llama.cpp code changes
Granite 4.2 ships three open models at 3B, 8B and 30B trained on 15T tokens
Granite 8B and 30B support OpenAI-compatible tool calling and 512K-token context
Cloudflare released Clef classification models for routing in agent pipelines

#IBM#Cloudflare#Granite#HuggingFace

Swift-Qwen3.8-27B Cuts Reasoning Tokens Without Sacrificing Much Accuracy

Security

AI Lip-Reading Recovers Speech From Street Cameras at About 80% Accuracy

Science

Quantized LLMs recover 97% accuracy; IBM adds reasoning to Granite

Swift-Qwen3.8-27B Cuts Reasoning Tokens Without Sacrificing Much Accuracy

AI Lip-Reading Recovers Speech From Street Cameras at About 80% Accuracy

IBM Research Proves Quantum Circuits Outperform LLMs on Two Problems

SoloQ's calibration-free quantizer cuts diffusion-LLM inference costs by 45%