October Local AI Roundup: Million-Token Agents on One GPU, 753B MoE in 273GB
vramkickedin · reddit · 2026-11-02
A Reddit roundup by vramkickedin collects local/open-source AI releases from October 2026. Highlights: Ornith-1.5-122B handles million-token agent tasks on a single GPU; KRAFTON's Raon-1-mini-preview runs Korean agentic AI on one RTX 5090; Flint is a MoE coding agent that investigates repos and fixes bugs on one GPU; Tencent's Hy-Memory-30B-A3B writes durable agent memory with 3B active params; LiquidAI's LFM2.5-24B-A2B targets fast on-device agent reasoning; an uncensored 3.0bpw quant fits a 753B MoE into 273GB; antirez's deepseek-v4.1-flash-gguf runs DeepSeek Flash on Apple Silicon; Intern-S2-397B scales scientific reasoning; LingBot-World 2.0 runs a world model at 16 FPS. On the multimodal side, MiniMax-M3.1-Flash offers 1M-token context. The month's trends: small active-param MoEs, single-GPU agent models, and long-context coders.
More from Infra
- Lumentum deal sparks debate: does the sale signal doubts about the AI cycle lasting? — BenBajarin · 2026-11-03
- Rohan Anil's CoreAuto trains MoE across two global clusters over a 25 Gbps link — _arohan_ · 2026-11-03
- Anthropic Q3: $17B revenue but swing to $567M operating loss, $895B compute commitments — rohanpaul_ai · 2026-11-03
- AI compute as a planet: three quarters under the US flag, China one seventh — LeviTurk · 2026-11-03
- Custom Qwen3.8-27B engine doubles llama.cpp speed on a 3090 with exact greedy outputs — Nice_Profession_9078 · 2026-11-02
- Uncensored DeepSeek-V4.1 quants cut refusals from 85% to 0% on 12 GB VRAM — gladkos · 2026-11-02