Alibaba made Qwen 3.8 Max generally available today (August 3): roughly 2.4 trillion parameters as a Mixture of Experts, a 1 million token context window, image and video input, and launch scores like 92.6 on GPQA Diamond and 86.6 on Terminal-Bench 2.1. The part that matters for this post: the open weight checkpoint is promised for the coming week, and it brings a consumer sized Qwen 3.8 27B along with it. Here is the practical picture for anyone who wants Qwen 3.8 on their own hardware.
The shape of the model decides everything
MoE models split the hardware question in two. Total parameters (2.4T) set your memory bill: every expert has to live somewhere. Active parameters (~95B per launch coverage) set your speed. That is why the Max serves cheaply in a datacenter, and also why it will never fit in your tower: all 2.4T parameters must sit in memory, because you never know which experts the next token activates.
The numbers for the Max
- FP8, the native serving precision: ~2.4 TB of weights.
- Q4, the local standard: ~1.2 TB.
- An extreme 2 bit quant: ~600 GB, with real quality loss.
The largest single machine a consumer can buy in 2026 holds 512 GB of unified memory, less than half of the Q4 weights, before you allocate a single byte of KV cache for that million token context. Someone will chain Mac Studios together for a single digit tokens per second demo within weeks of the weights dropping. It will be a great video and a bad daily driver.
The 27B is the actual local story
This is the difference between this launch and Kimi K3 or DeepSeek V4 Pro: the same drop includes an open weight 27B. Projected from the Qwen 3.6 27B precedent:
| VRAM | Quant class | Precedent size |
|---|---|---|
| 8 GB | 2 bit (UD-IQ2 class) | 8.7 GB, reduced quality |
| 12 GB | Q3_K_M | 13 GB, RTX 3060 sweet spot |
| 16 GB | Q4_K_M, recommended | ~16 GB |
| 24 GB | Q6_K, near lossless | ~21 GB |
Community GGUFs of open Qwen releases usually appear within days of the weights. If the 27B inherits even part of the Max's agentic gains (FrontierSWE jumped from 40.7 to 73.5 this generation), it becomes the default local model in its class more or less immediately. And weights mean derivatives: every open Qwen generation has received abliterated and heretic builds within weeks.
What to run today, not next week
Until the drop, the newest Qwen you can actually run is Qwen 3.6, and it is genuinely strong: the 27B dense runs from 8 GB VRAM up, the 35B MoE (3B active) is the coding pick at 24 GB. One line with a current llama.cpp era stack:
ollama pull qwen3.6:27b
If you would rather click than type, the free and open source local AI studio Locally Uncensored has the Qwen family in its one click model catalog, checks your memory before you download, and speaks to 12 local backends. The full setup walkthrough with quant tables lives here: How to Run Qwen 3.6 Locally, and the complete Qwen 3.8 hardware math is in Can You Run Qwen 3.8 Locally?
Why bother running it at home
Privacy: your prompts never leave the machine. Cost: open weights turn a metered bill into a one time hardware decision. Control: the model you benchmark today is the model you run next year, no silent upstream swaps. For the first time, the frontier launch everyone is hyping comes with a version of itself you will actually own, in the same week's drop.
Top comments (0)