Community asks: where is the Qwen 4 35B MoE as the Qwen 4 lineup nears
solyarisoftware · x · 2026-10-31
- Community reports the Qwen 4 season is approaching, with Max, Flash, Plus and a 27B dense model expected; the 27B is rumored to be a powerhouse, though whether it fits a single 24GB RTX 3090 remains unclear.
- @analogalok recalls why Qwen 3.6 35B A3B mattered: not chart-topping, but running well on a 6GB secondhand card — architectural efficiency that delivered daily-driver reasoning on a $250 GPU.
- He asks the Qwen team for a 35B-class MoE in Qwen 4: "once developers get real intelligence on cheap hardware, they never forget what that freedom felt like."
More from Infra
- RTX 4090 BF16 unlock via UEFI toolbench: ~10% faster image gen, still trails RTX 5090 — devmoching · 2026-10-31
- Running a Plan-then-Execute Local Agent on 12GB VRAM: Big Model Plans, Small Model Acts — Loose_Doubt367 · 2026-10-31
- PhobosStream: Run 14B Video DiTs on 8GB GPUs With 788MB Peak VRAM via Disk Streaming — martianacademy · 2026-10-31
- Over $100M in Data-Center Equipment Stolen Since June as GPS Spoofing Emerges — BuddyWeary653 · 2026-10-31
- Musk predicts orbital compute will outpace Earth's by end of 2028 as SpaceX targets 1M tons to orbit — XFreeze · 2026-10-31
- Lawyer runs DeepSeek-V4.1-Flash on 2x DGX Spark + 3090s: 2,618 tok/s prefill at 256K, 95.9% KLD accuracy, all data open — TumbleweedNew6515 · 2026-10-31