TL;DR: Qwen3.8-Max (the 2.4T-parameter checkpoint Qwen/Qwen3.8-2.4T-A95B, on Hugging Face since August 12, 2026) is downloadable, but it is not Apache 2.0 — it ships under a bespoke “Qwen3.8-Max License” with revenue-triggered terms. At 4-bit it still occupies roughly 1.3–1.5 TiB of weights, which means an 8-GPU data-center node at minimum. For almost everyone, the model to actually self-host is Qwen3.8-27B, which is genuinely Apache 2.0 and runs on one 24 GB card.

	Qwen3.8-Max open weights	Qwen3.8-27B	Qwen3.8-Max API
Best for	Labs with 8+ data-center GPUs; fine-tuning rights	Everyone self-hosting at home or on one server	Max-class quality without hardware
License	Qwen3.8-Max License (custom, revenue triggers)	Apache 2.0	Alibaba Cloud ToS
Hardware / cost	~1.3–1.5 TiB VRAM at 4-bit (8× B300-class)	~17 GB VRAM at Q4_K_M	$2 in / $6 out per M tokens (Aug 2026)
The catch	Not Apache 2.0; text-only checkpoint	Far below Max on hard benchmarks	Your prompts leave your network

Honest take: Download the 27B, rent before you ever buy for the big one, and treat Qwen3.8-Max’s open weights as what they are — a fine-tuning asset for funded teams and price pressure on the API market, not something you will run in a homelab.

Is Qwen3.8-Max actually Apache 2.0?

No. The checkpoint Alibaba published at Qwen/Qwen3.8-2.4T-A95B carries a custom “Qwen3.8-Max License”, not Apache 2.0 — whatever the launch-week Reddit threads said. The license file on the Hugging Face repo (checked October 2026) sets out three terms that matter:

• $50M MaaS trigger. If you (or an affiliate) run a Model-as-a-Service or “AI Work Assistant” business with aggregate revenue above US$50 million in any consecutive twelve months, you must obtain a separate commercial license from Qwen before using the model commercially.
• Attribution above scale. Products with more than 100 million monthly active users or US$20 million in monthly revenue must display the model name prominently in the UI.
• Internal-use carve-out. The separate-license requirement does not apply to internal use, as long as the model, its outputs, or its capabilities are not offered to third parties.

A claim circulated in August 2026 that the license blocked downloads from the US, EU, UK, and South Korea. The published license text contains no territory clause — that one appears to have been confusion with MiniMax’s H3 Community License, which really does exclude those regions.

Two practical consequences. First, for a homelab or a company under those thresholds, the license is permissive in effect: you can download, run, fine-tune, and use outputs commercially. Second, it is still not OSI-approved open source, and the terms can matter if you’re building a product you hope will grow — the same class of catch we documented for GLM-5.3 and other 2026 releases in open-weight licenses that block commercial use. The variant that is clean Apache 2.0 is Qwen3.8-27B — setup covered in our Qwen3.8-27B self-hosting guide.

One more asymmetry worth knowing before you download 4+ TiB: the open checkpoint is text-only. The hosted Qwen3.8-Max API accepts images and video; the weights on Hugging Face do not.

How much VRAM does Qwen3.8-Max need?

About 4.45 TiB at BF16, and roughly 1.3–1.5 TiB for the 4-bit (FP4/AWQ) variants — weights alone, before KV cache. Qwen3.8-Max is a mixture-of-experts model: 2.4 trillion total parameters with about 95 billion active per token. MoE routing cuts compute and per-token bandwidth, but it does not cut memory — every expert has to sit in VRAM waiting to be routed to.

Precision	Approx. weights on disk/VRAM	Fits on
BF16	~4.45 TiB	32+ × 141 GB-class GPUs
FP8	~2.2 TiB	16× B300-class or 24× H200
FP4 / 4-bit AWQ	~1.3–1.5 TiB	8× B300 / 8× MI355X; H200 needs 16+

The recipe Alibaba and the vLLM community converged on in the first weeks: FP4 weights on a single 8-GPU node of NVIDIA B300s or AMD MI355Xs, FP8 KV cache, tensor parallelism across the node. On H200s (141 GB each), eight cards give you 1.13 TiB — under the 4-bit weight footprint, so H200 deployments need two nodes and pipeline parallelism, with the interconnect pain that brings.

For speed intuition, the decode ceiling is bandwidth ÷ bytes-read-per-token. At 4-bit, the ~95B active parameters mean roughly 50–55 GB read per token. That’s why nobody serious attempts this on CPU-offload rigs: a dual-channel DDR5 desktop at ~90 GB/s tops out below 2 tok/s even if you could attach 1.5 TiB of RAM — which, at October 2026 DDR5 prices, would cost more than the GPUs.

Can you run Qwen3.8-2.4T-A95B on consumer hardware?

No — and it fails fast, which is at least diagnostic. vLLM (which added Qwen3.8 architecture support at launch) will happily start loading on whatever you have:

vllm serve Qwen/Qwen3.8-2.4T-A95B \
  --tensor-parallel-size 8 \
  --max-model-len 32768

On anything below the ~1.4 TiB tier, the load dies partway through shard initialization:

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.92 GiB
(GPU 0; 23.99 GiB total capacity; 22.41 GiB already allocated)

A real gotcha we hit reading the deployment configs rather than the headline specs: the model card advertises 262K native context (1M via YaRN), but the official vLLM launch command ships with --max-model-len 32768. That’s not a typo — KV cache for long contexts on a model this wide eats tens of extra gigabytes per request, and the reference 8-GPU FP4 node has almost nothing left after weights. If your plan assumed “262K context, it’s on the card,” budget GPUs for the KV cache separately or run short contexts. The fix for long-context work on this model is more hardware, not a flag.

llama.cpp --tensor-split across consumer cards doesn’t rescue you either: sixty-plus RTX 3090s’ worth of VRAM across PCIe is not a serving plan. In the first two weeks post-release, community GGUF quants existed but every credible multi-GPU report came from H200/B300-class clusters.

What does it cost to run vs. renting or the API?

Renting the minimum viable cluster costs more per day than most people’s monthly API bill. Marketplace H200s run from about $2.20/GPU-hour on Vast.ai (floating, October 2026), and you need ~16 of them — from roughly $35/hour, or ~$850/day if they’re actually available in one fabric. Buying the 8× B300 node is well into six figures.

The hosted Qwen3.8-Max API was priced at $2 in / $6 out per million tokens at launch (August 2026). At $850/day, that’s the equivalent of ~140M output tokens daily — a volume almost no self-hoster reaches. The arithmetic only flips if you’re saturating the cluster around the clock or you need fine-tuning and data isolation that no API offers; the general framework is in rent vs own GPU math for open models.

When NOT to self-host Qwen3.8-Max

Which is to say: almost always. Skip it if any of these apply —

• You have fewer than 8 data-center GPUs. There is no consumer path. Not quantization, not offload, not a 3090 farm.
• You wanted the multimodal Max. The open checkpoint is text-only; image/video input is API-exclusive.
• You’re building a product you expect to exceed $50M MaaS revenue. You’ll need a separate Qwen license anyway — price that in now, not at due-diligence time.
• Your actual need is “a strong local model.” Qwen3.8-27B at Q4_K_M (~17 GB) on a used RTX 3090 covers that for about $1,400 of hardware.

The genuine self-hosting cases: labs fine-tuning on proprietary data (the license permits it below the triggers), regulated environments where prompts cannot leave the building, and infrastructure teams hedging against API repricing.

What to actually buy

Prices as of October 2026, all verified above or in the linked guides:

Your situation	The move	Price	Where
You want Qwen3.8-class output at home	Used RTX 3090 24GB + Qwen3.8-27B	~$1,399–$1,450	Check price
You want to test the 2.4T model before committing	Rent H200s by the hour	from ~$2.20/GPU·hr	Vast.ai
You need Max quality occasionally, no infra	Qwen3.8-Max API	$2 in / $6 out per M tokens	Alibaba Cloud Model Studio

FAQ

Is Qwen3.8-Max open source? It’s open-weight, not open source. The Qwen3.8-Max License allows free download, self-hosting, fine-tuning, and commercial use below revenue thresholds ($50M/12mo for MaaS businesses), but it is not an OSI-approved license. Qwen3.8-27B is the Apache 2.0 release in the family.

How much VRAM does Qwen3.8-Max need at 4-bit? Roughly 1.3–1.5 TiB for weights alone, plus KV cache. The reference deployment is one 8-GPU node of NVIDIA B300s or AMD MI355Xs; 141 GB H200s need 16+ cards across two nodes.

What’s the cheapest way to try Qwen3.8-Max yourself? The API, by a wide margin ($2/$6 per M tokens). If you specifically need to run the weights — say, to validate a fine-tune — rent marketplace H200s from ~$2.20/GPU-hour on Vast.ai rather than buying anything.

Sources

• Qwen/Qwen3.8-2.4T-A95B LICENSE file, Hugging Face — license terms, checked October 2026
• Qwen3.8 open weights ship under two different licenses — SQ Magazine
• Qwen3.8-Max open weights: first Max-class Qwen you can download — llm-stats.com
• How to deploy Qwen3.8-Max locally — CometAPI — BF16/FP4 footprints, reference vLLM recipe
• Hacker News discussion of the Qwen3.8-2.4T release

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe