How a 100B+ Model Fits in a Gaming PC — what Strata and Project Maya changed

Local AI Updated 2026-11-04 · 3 min read · #moe #quantization #ram #ssd #engines

The one trick behind the whole trend

A 125–748B MoE model does not need 125–748B of fast memory. It needs:

All experts stored compressed (2–4-bit packs) — the bulk lives in system RAM;
A big lookup/n-gram table on NVMe SSD (≈29 GB) — slow but cheap;
Your GPU only ever sees the active slice per token — a 12 GB card is suddenly a datacenter card;
Draft layers (MTP / DSpark / DFlash2) speed up decoding because active params × memory bandwidth, not total size, decides tok/s.

That’s it. A used 3060 + 64 GB RAM box runs the biggest models at “slow typist” speeds; a fast 12 GB card + 64 GB RAM runs the trend’s face at 76–94 tok/s.

What actually runs at home (Nov 2026, sizes are weights in RAM+VRAM after quant)

Model	Class	Smallest realistic pack	What it takes
Qwen3.8-Flash-Next (125B)	the trend’s face — chat/code/vision	Q2_0 ≈ 37.6 GB	48 GB RAM + any recent GPU; 64 GB RAM fits every size
GLM-5.3-Flash (320B, 18B active)	the agentic/code flagship	Maya-XS ≈ 73.4 GB	64 GB RAM + 24 GB card at 128K ctx
MiniMax M3.1-Flash (428B, 23B active)	newest (open-sourced Oct 23) — most raw capacity	Maya-XS ≈ 93 GB	a 128 GB box to hold the whole model in memory
DeepSeek-V4.1-Flash-1017 (748B)	the biggest — the “748B on a gaming PC” headlines	1.75 bpw ≈ 239 GB on disk	46–64 GB RAM + 12 GB+ GPU + huge NVMe; 6–9 tok/s on a 3060, 55–69 with a drafter on 4× R9700
Qwen “Coder” variant (half the experts kept)	code-only budget builds	IQ1_M ≈ 32 GB	32 GB RAM + 12–24 GB GPU

Which engine? (this is the real decision now)

Project Maya — a three-tier scheduler: the experts your conversation uses most go on the GPU, the next ones in RAM, the rest on NVMe, and it moves them as you chat. Runs GLM-5.3-Flash, MiniMax-M3.1-Flash and DeepSeek-V4.1-Flash on one NVIDIA card (up to 16). Community reports say its draft-layer handling beats ExLlamaV3 where other paths actually slow down.
Strata — the friendly guided launcher: strata launch --plan prints exactly what fits your machine before anything downloads. Home of the Qwen3.8-Flash-Next packs (ISTA-DAS sizes, Coder, Unsloth 4–8-bit, NVFP4, Gyro).

The two bills, unchanged

Context is a second bill: every number above was measured at a specific context (8k for the 748B class, 32–128K for the Flash-Next class); bigger context = more GB. Check before buying.
Trade headlines: every hot model here is China-origin; the export-catalogue draft (Oct 2026, comment period) could add paperwork for overseas commercial use — weights stay on HuggingFace worldwide today, and we track per-model status on each page.

📌 Summarised from official docs and hands-on samples. Product features and pricing change often — official site wins.

← Newer: GLM-5.3-Flash: the 320B That Fits Because Its Attention Does Not Grow Older: MiniMax M3.1-Flash: the 428B 'Space Bunny' Everyone Switched To in October →