multimodal

machine learning Aug 14, 2026 7 min read

Qwen3.8-27B FP8: FP8 Quantization, Thinking Control, and 1M Context in Practice

Qwen3.8-27B-FP8 packages a 27B vision-language model as fine-grained FP8 (block size 128) weights. It adds practical controls like `reasoning_effort` and `preserve_thinking`, plus native 262K context with extensibility toward 1M tokens for long-horizon tasks.

by ahsan

#inference #llm #machine learning #multimodal #quantization

machine learning Jul 15, 2026 6 min read

Inkling (975B): What Open-Weights Mixture-of-Experts Really Means—and How to Fine-Tune It

Thinking Machines’ Inkling is an open-weights, multimodal MoE LLM with 975B total parameters and 41B active per token. It supports long context (up to 1M tokens), controllable thinking effort, and fine-tuning via Tinker.

by ahsan

#fine-tuning #llm #mixture of experts #multimodal #open-weights

ai systems & inference Jul 15, 2026 7 min read

Bonsai 27B on a Phone: How 1‑bit & Ternary Quantization Make Agentic AI Possible

Bonsai 27B is a 27B-class multimodal model announced by PrismML that they claim can run on a phone by using 1-bit or ternary weight formats with group-wise FP16 scaling. The architecture also includes on-device vision support and speculative decoding to make long agentic loops practical without shipping user data to the cloud.

by ahsan

#agentic systems #llm quantization #multimodal #on-device ai #speculative decoding