Qwen3.8-Flash-Next: the 125B That Made 100B-at-Home a Hobby

Local AI Updated 2026-11-04 · 3 min read · #qwen #moe #multimodal #local

Status board (checked 2026-11-04)


Who made it	Qwen team (Alibaba)
License	Qwen Community License — check the model card for commercial thresholds
Export-control status	⚠️ China catalogue draft (comment period) — watch the date on this box

The numbers


Total params	~125B MoE; every pack reads pictures too
How it fits	shard 1’s experts load into RAM, a 29 GB lookup table stays on SSD, your GPU only ever handles the active slice
Fit rule	RAM ≥ the size’s experts + ~10 GB for the OS. A bigger GPU makes it faster, it does not lower the RAM needed
Packs (RAM+VRAM)	Q2_0 ≈ 37.6 GB · IQ2_XS ≈ 39.2 · IQ3_XXS ≈ 47 · IQ3_S ≈ 54.8 — “matches the full model on the published tests”
Measured speed (RTX 5070 12 GB, 64 GB RAM)	Q2_0: 94 tok/s short chat, 76 at 128K ctx; IQ2_XS 79/63; IQ3_XXS 62/49; IQ3_S 53/46. A 32K prompt reads at ~2,650 tok/s

Getting started (3 steps)

Install your engine (Strata is the guided one) and run its plan command — it shows exactly what fits before anything downloads (lower sizes are a 66–76 GB download; the MTP draft layer adds ~6 GB).
Pick by RAM: 64 GB fits every size (IQ3_S with little else open); 48 GB fits Q2_0/IQ2_XS; with 32 GB RAM + a 24 GB card a low-RAM mode maps experts from files instead.
Vision is on by default: the picture reader runs on CPU under 24 GB of VRAM (a few seconds per image, no VRAM used) and on the card from 24 GB up.

Best use

The one-model-for-everything pick: general chat, decent code, and picture reading, at speeds that feel real rather than “slow typist.”

Variant	When it’s the right call
Coder (IQ1_M, keeps 256 of 512 experts per layer, ~30 GB)	32 GB RAM boxes; its authors measured 91% of the full model’s SWE-bench Verified score, 99% of LiveCodeBench — but weaker outside code
Unsloth’s 4–8-bit files	72–136 GB rigs that want every expert in RAM at 4-bit+
Gyro-S (27.6 GB)	the smallest option of all, lower quality, never the default

Honest weaknesses

The quality ladder is real: Q2_0 is “good”, IQ3_S is what matches the full model on published tests — and IQ3_S wants 54.8 GB of your RAM.
On 48 GB boxes you’re on Q2_0/IQ2_XS and every other program competes for what’s left.
Chinese-first ecosystem: the best numbers live in the engine’s docs, not in one-click defaults.

FAQ

What’s the “Uncensored” build I keep hearing about?

OrcaRouter’s Uncensored IQ3_XXS is a community repack that removes refusal behavior — it needs an explicit packing conversion step, not an installer menu item. (The model drafting this very page is exactly that build, Q4_K_M.)

Is there a smaller sibling?

Yes — Qwen3.6-35B-A3B (35B total / 3B active, Apache 2.0, 23–38 GB across three sizes) is the “my PC is a potato” entry drug: 59.7 tok/s on an RTX 3060 with experts on CPU.

Does the export-catalogue story touch it?

Same as the other Chinese models: weights stay on HuggingFace worldwide today, the draft adds paperwork noise rather than a wall — and this page’s status box gets a dated line when anything changes.

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

← Newer: Project Maya: the Scheduler That Puts 748B on Your Desk Older: Strata: the Engine That Made 125B a Gaming-PC Hobby →