← Dashboard  /  News

2026-10-09

AI/Tech Digest 2026-10-09

AI & local LLM

1. $2800 rig with 8x Radeon Pro V620 (256 GB VRAM) + custom vLLM fork = Qwen3.8-Flash-Next at 60 to 100 t/s decode and 3000+ t/s prefill [LAB]
   Demonstrates high-throughput local inference performance using consumer-adjacent hardware and custom serving forks.
   https://www.reddit.com/r/LocalLLaMA/comments/1x0wnz1/2800_rig_with_8x_radeon_pro_v620_256_gb_vram/ (reddit.com)
2. Saluki 27B: "96% of Qwen 3.8’s performance at ~1/7 the size" [LAB]
   Relevant for the reader's interest in mid-size models and efficient architecture for local/edge inference.
   https://www.reddit.com/r/LocalLLaMA/comments/1x0mn7x/saluki_27b_96_of_qwen_38s_performance_at_17_the/ (reddit.com)
3. Introducing Mistral Large 4 [LAB]
   A major frontier model release from a key player in the open-weight/open-access ecosystem.
   https://mistral.ai/news/mistral-large-4/ (mistral.ai)