Reddit asks: which 50-150B MoE models are worth running locally besides Qwen Flash Next
FlyingFishMakeAWish · reddit · 2026-10-22
The poster upgraded desktop RAM to run 50-150B MoE models split across RAM and VRAM, and asks for picks beyond Qwen Flash Next.
Shortlist: GPT-OSS-120B (still usable?), Mistral Small 119B, Qwen3.5 122B A10B (supplanted by Flash Next?), and Ling 3.0 Flash. Agentic coding is explicitly out of scope — Flash Next already rules there and they use Claude/Codex for code — so they want models that excel elsewhere, including abliterated or community fine-tunes.
More from Infra
- SpaceX's Bastrop site balloons into industrial city with 1.1M sq ft solar factory — XFreeze · 2026-10-22
- How to benchmark compiler opts: inspect emitted code, split cold vs steady — cfregly · 2026-10-22
- Nested CUDA graphs cut replay launches from 16 to 2 per call — cfregly · 2026-10-22
- Nested graphs deliver 4.5× steady-state speedup in depth-8 benchmark — cfregly · 2026-10-22
- Open-source tool condenses millions of logs into an LLM-sized snapshot — CommanderAlPachino · 2026-10-22
- Anthropic accounts for $90B of $157B neocloud deals, powering the GPU rental boom — FinanceYF5 · 2026-10-22