TL;DR: Alibaba previewed Qwen3.8-Max on July 19, 2026 — 2.4 trillion total parameters, multimodal, roughly 1M-token context, and a claim that it’s “second only to Fable 5.” What Alibaba did not publish: active-parameter count, benchmarks, weights, a license, or a release date. Nothing in a home lab runs it today, and you can’t even plan a build yet.

	Qwen3.8-Max-Preview (cloud)	Wait for open weights	Run Qwen3.6 locally today
Best for	Curiosity, cheap evaluation	Cluster owners, patient builders	99% of home labs
Price / Cost	From $6 (Token Plan Lite tier, preview at 10% of standard)	Unknown — est. 840GB+ memory even at 2-bit	Used RTX 3090 ~$1,050–$1,252
The catch	Credit-based, no per-token rate, prompts leave your machine	No date, no license, no active-param count published	35B-A3B, not 2.4T — but it’s yours

Honest take: Don’t spend a dollar of hardware budget on Qwen3.8-Max. Alibaba has published exactly one number — 2.4T total parameters — and that number alone guarantees no consumer machine will ever hold it. Try the preview through the Token Plan if you’re curious, keep running Qwen3.6-35B-A3B on your 24GB card, and watch what Kimi K3’s July 27 weight drop does to this segment first.

What Alibaba actually announced

On July 19, 2026 — at the World Artificial Intelligence Conference in Shanghai, and three days after Moonshot AI’s Kimi K3 launch — the Qwen team posted that “Qwen3.8 is launching and going open-weight soon,” and shipped Qwen3.8-Max-Preview to Alibaba’s Token Plan subscription, Qoder, and QoderWork the same day.

The confirmed facts, from Alibaba’s own announcement and the Qwen Cloud integration metadata:

• 2.4 trillion total parameters — the first Qwen multimodal model above 1T
• Multimodal: text, images, video, and documents in one model
• 983,616-token context window with a 131,072-token maximum output (the “1M context” in coverage rounds up from that metadata figure)
• Preview availability now, at 10% of standard pricing during the preview period
• Open weights “soon” — Alibaba’s word, with no date attached

And the vendor claim doing all the marketing work: Qwen3.8 is “one of the most powerful models available today, compatible to leading frontier AI models, second only to Fable 5.”

That claim shipped with no scores, no named test suite, and no methodology. As of July 24, the only independent head-to-head we can find is TrilogyAI’s StackPerf architecture benchmark, which scored Kimi K3 at 83 and Qwen3.8 at 80 — one run, on one task. Treat it as a data point, not a verdict.

What Alibaba didn’t publish (and why it blocks every VRAM calculation)

When Alibaba launched Qwen3.7-Max in May, it came with a model card. Qwen3.8-Max-Preview came with a total-parameter count and a tweet. Still unpublished as of July 24, 2026:

• Active parameters per token. Coverage describes a sparse Mixture-of-Experts, but Alibaba hasn’t confirmed the architecture, the expert count, the routing, or how many parameters fire per forward pass.
• Benchmark results. Not one number.
• License. “Open-weight soon” could mean Apache 2.0 (like Qwen3.6) or something more restrictive.
• A release date for the weights.

Here’s why the missing active-parameter count matters more than anything else on that list. In a sparse MoE, total parameters set your memory bill; active parameters set your speed. Kimi K3 runs 16 of 896 experts per token; Inkling runs 41B active out of 975B total. Those two numbers are what let us compute “you need X GB and you’ll get roughly Y tok/s.” With Qwen3.8-Max we have half of one equation: we know the memory wall is enormous, and we know nothing about the speed you’d get for climbing it.

This is the same discipline we applied to Kimi K3 before its numbers firmed up: when the vendor hasn’t published the spec, we won’t invent it.

The memory math you can actually do

One thing needs no disclosure to compute: 2.4 trillion parameters have a physical size, and quantization can only shrink it so far. We won’t publish fake GGUF sizes — instead, here are the confirmed, on-disk build sizes of the closest comparable models, all verified in our earlier guides:

Model (confirmed sizes)	Total params	Full-precision release	Dynamic 2-bit build
Kimi K2.6	~1T	610GB	350GB
Kimi K2.7 Code	~1T	605GB	325GB (UD-Q2_K_XL)
Qwen3-Coder 480B	480B	~960GB BF16	~180GB (UD-Q2_K_XL)
Kimi K3	2.8T	~1.4TB (native MXFP4)	TBD (weights due Jul 27)

Scale those to 2.4T and you get the only honest Qwen3.8 estimates possible today — flag: every number in this table is our estimate, not a published figure:

Qwen3.8 build (estimated)	Approx. size	What could hold it
4-bit (if shipped MXFP4-style like K3)	~1.2TB	Nothing consumer. Not 8× H100 (640GB). Multi-node datacenter or a maxed Mac Studio cluster, barely
Dynamic 2-bit GGUF (scaled from K2.6’s 0.35 bytes/param)	~840GB	A 4× Mac Studio RDMA cluster (~$40K, 1.5TB pooled) with room for KV cache — at a real quality cost
Anything a 24GB card fits	—	Does not exist and never will for this model

For perspective: the ~$40K four-Studio cluster we covered in July runs the 1T Kimi K2 at about 25 tok/s. Qwen3.8-Max is 2.4× larger in total parameters, and until Alibaba publishes the active-parameter count, nobody can tell you whether it would decode at a usable speed even on hardware that holds it. That uncertainty — not the price tag — is the real reason to wait.

Qwen3.8-Max vs. Kimi K3 vs. Inkling: the trillion-class scoreboard

Three “frontier-class open” announcements in ten days, in very different states of actually being open:

	Qwen3.8-Max	Kimi K3	Inkling
Total / active params	2.4T / undisclosed	2.8T / ~40–60B est. (16-of-896)	975B / 41B
Context	983,616 tokens	1M	1M
Weights	”Soon,” no date	Due July 27, 2026	Released July 15 (Apache 2.0)
License	Unannounced	Modified MIT	Apache 2.0
Verified benchmarks	None published	AA Intelligence Index 57 (verified across nine evals)	AA Index 41, SWE-bench Verified 77.6%
API price	Credit-based only, no per-token rate	$0.30 cached / $3 in / $15 out per M	Tinker, ~$1.87/$4.68 per M at launch

The pattern is hard to miss: Kimi K3 published its architecture, benchmarks, license, and a weight date. Inkling shipped its weights on day one. Qwen3.8-Max published a parameter count and a ranking claim. Alibaba has earned benefit of the doubt — Qwen3.6 is Apache 2.0 and the most-recommended 24GB model on this site — but until the weights and the spec sheet land, this is a cloud product with a press release attached.

Trying the preview without getting burned

There’s no standalone per-token API price — unusual for a Max-tier model. Access runs through Alibaba’s Token Plan, a credit-based subscription: the Lite tier is $6 for 2,500 credits per 7 days, scaling to a $68 Pro tier with 40,000 credits and 6–8 concurrent agents. During the preview, Qwen3.8-Max is billed at 10% of standard pricing, with an additional heavy off-peak discount during night hours, and it’s also wired into Alibaba’s Qoder and QoderWork coding products.

Two practical cautions before you sign up:

1. Credit burn is unpredictable. With no published input/output rate, you can’t forecast cost per task the way you can with Kimi K3’s $3/$15. Budget the $6 tier as a toy, not a pipeline.
2. Preview pricing is promotional. The 10% number is temporary by definition. Don’t build a workflow around a rate that will 10× when the preview ends.

If you want to wire a trillion-class model into a coding agent today with per-token pricing you can actually reason about, K3’s API or Inkling on Tinker are the saner picks — see our sister site aicoderscope.com for wiring cloud and local backends into Cursor and Cline.

What to run on your own GPU while you wait

The queue of “2.4T model announced → home-labbers ask what GPU to buy” has the same answer it had for Kimi K3, Inkling, and Qwen3-Coder 480B: your money buys more capability in the 24GB class than in a deposit toward a model you can’t hold.

A used RTX 3090 is running $1,252 market average across 336 listings in July 2026 (ResalePrices), with eBay lows around $1,050 (BestValueGPU) — still the bandwidth king of the used market at 936 GB/s. On that card, Qwen’s own Qwen3.6-35B-A3B — Apache 2.0, ~22GB at Q4_K_M — decodes at 50–65 tok/s, which is the difference between an assistant and a spectator:

$ ollama pull qwen3.6:35b
$ ollama run qwen3.6:35b --verbose
>>> Explain PCIe ACS and why it matters for dual-GPU inference.
...
eval rate:            57.31 tokens/s

(The 57 tok/s line is an illustrative mid-band reading — your exact number will vary with context length and quant.) Anything in that 50–65 tok/s band means your GPU is doing what the 2.4T preview can’t: running on hardware you own, offline, with no credit meter. If your workload genuinely needs frontier-class quality, rent it by the hour — a RunPod H100 is $2.89/hr and an A100 80GB is $1.39/hr (referral link) — and keep the local stack for the 80% of tasks a 35B MoE handles. Our rent-vs-buy math hasn’t changed.

For the capacity-first crowd, a Mac Studio M3 Ultra or a used-3090 pair remains the realistic ceiling — our 70B-on-24GB guide covers what partial offload actually feels like.

When the weights land: the two day-one traps

Assume Alibaba does ship Qwen3.8 open weights. Two failure modes will hit early adopters in the first 48 hours, because they hit every big MoE launch this year:

Trap 1: Error: unknown model architecture. Day-one community GGUFs routinely land before runtime support. In June, Ollama 0.30.7 users pulling DiffusionGemma GGUFs got exactly unknown model architecture: 'diffusion-gemma' (Ollama issue #16664) until a runtime update shipped. A new Qwen attention or routing variant would do the same. The fix is never re-downloading — it’s updating Ollama/llama.cpp, or waiting for the official ollama pull tag. Full triage in our unknown-architecture fix guide.

Trap 2: the pre-flight memory refusal. Even a 2-bit ~840GB build will make Ollama refuse to load with model requires more system memory (X GiB) than is available on anything short of a specced-out server — sparse activation saves bandwidth, but every expert must stay resident. We’ve seen readers burn a day downloading a 300GB+ model their box could never load; the memory-check fix guide explains what the numbers mean. Check the math before the download.

Verdict: a press release is not a hardware target

Qwen3.8-Max is probably a very strong model — Alibaba’s track record with the Qwen3.6 line earned that presumption. But a hardware guide runs on numbers, and right now the entire public spec sheet is: 2.4T parameters, ~1M context, multimodal, “second only to Fable 5,” trust us.

What to watch, in order of how much it would change this article:

1. The active-parameter count. This decides whether a CPU-offload or Mac-cluster build decodes at 5 tok/s or 25 tok/s.
2. The license. Apache 2.0 (like Qwen3.6) makes it a serious self-hosting target; anything restrictive makes it a curiosity.
3. Unsloth’s GGUF sizes when weights land — the moment estimates in this article get replaced with measurements.
4. Kimi K3’s July 27 weight release — the first 2.8T open weights will set the community’s expectations (and tooling) for everything in this class, Qwen3.8 included.

We’ll update the moment any of those four land. Until then: the preview costs $6 to satisfy your curiosity, and your GPU budget is better spent on hardware the 2026 open-model leaders actually fit. For the FOSS-side view of what “open-weight soon” should mean — license terms, reproducibility, what counts as actually open — our sister site aifoss.dev covers the open-model ecosystem in depth.

FAQ

Can I run Qwen3.8-Max locally? No. The weights aren’t released, and at 2.4T total parameters even an estimated 2-bit quant (~840GB) exceeds any consumer machine. A ~$40K four-Mac-Studio cluster is the only home-adjacent hardware that could plausibly hold it, and decode speed is unknowable until Alibaba publishes the active-parameter count.

Is Qwen3.8-Max a Mixture-of-Experts model? Coverage describes it as a sparse MoE, but Alibaba has not officially confirmed the architecture, expert count, or active parameters. That’s the single most important undisclosed number for anyone pricing hardware.

When do the open weights come out? Alibaba says “soon” with no date, license, or model card. By comparison, Kimi K3 (announced three days earlier) committed to July 27, 2026 under Modified MIT.

Is it really the second-best model in the world behind Fable 5? That’s Alibaba’s claim, published with no benchmarks or methodology. The only independent test so far (TrilogyAI’s StackPerf, single run) put it slightly behind Kimi K3, 80 vs 83. Verified rankings will have to wait for Artificial Analysis and the standard suites.

What’s the best Qwen model I can actually run today? Qwen3.6-35B-A3B: Apache 2.0, ~22GB at Q4_K_M, 50–65 tok/s on a used RTX 3090 ($1,050–$1,252, July 2026). See the full guide, or Qwen3.6-27B if you prefer dense.

Recommended Gear

Products linked in this article:

• RTX 3090 (used) — 24GB, 936 GB/s; runs Qwen3.6-35B-A3B at 50–65 tok/s
• Mac Studio M3 Ultra — the unified-memory capacity path for 100B+ models

Sources

• Qwen3.8 launch announcement — Alibaba Qwen on X
• Alibaba’s Qwen takes on Kimi K3 with open-weight Qwen 3.8, says model is “second only to Fable 5” — The Decoder
• Alibaba Previews Qwen3.8-Max, a 2.4 Trillion-Parameter Multimodal Model — MarkTechPost
• Alibaba unveils Qwen3.8-Max Preview with 2.4T parameters; open weights coming soon — FoneArena
• Qwen3.8 Max Preview pricing, providers, and specs — Models.dev
• Qwen3.8-Max pricing: the preview deal and hidden costs — eesel AI
• Qwen 3.8 Max Benchmark: How It Compares With Kimi K3 — TrilogyAI
• Qwen 3.8 Max vs Kimi K3: What We Can Compare Today — Emergent
• RTX 3090 Used GPU Price & Fair Asking Range — ResalePrices
• RTX 3090 Price Tracker US, July 2026 — Best Value GPU

Last updated July 24, 2026. Prices, model availability, and specs change quickly; verify current rates before purchasing.