Qwen3.8-Max: Verified Specs, API Pricing, Benchmarks, and Local Mac Limits

Historical snapshot: August 3, 2026. Qwen3.8-Max is Alibaba’s new flagship Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion active parameters per token. QwenCloud lists a 1-million-token context window, $2 per million input tokens, and $6 per million output tokens. The weights were promised for the following week but were not yet available at that snapshot. A full 2.4T checkpoint would not fit in the 512 GB unified-memory ceiling of today’s largest Mac, even at a theoretical 2-bit raw weight size.

Update: August 25, 2026. The official Qwen organization now lists the Qwen3.8-27B checkpoint, the full Qwen3.8-2.4T-A95B checkpoint, and an official FP8 variant. The Max-class open weights are therefore public. The open 2.4T checkpoint is text-only, requires thinking, and does not support multimodal input; the hosted Qwen3.8-Max version adds vision, non-thinking, and built-in tools. Its custom license has additional conditions. The cloud figures below remain the historical August 3 snapshot; current regional Model Studio prices are added immediately before the pricing table.

Verified snapshot

Item	Status on August 3, 2026
Official product name	Qwen3.8-Max
Official Alibaba model ID	qwen3.8-max
Architecture	Mixture of Experts
Parameters	2.4T total, 95B active per token
Inputs / output	text, image, video / text
Context	1M; 991K max standard input; 131K max output
Hosted price	$2 input / $6 output per 1M tokens
Open weights	released: 2.4T-A95B and FP8; custom Qwen3.8-Max license
Single-Mac deployment	not practical for the full checkpoint

Why the “95B active” number needs context

Qwen3.8-Max routes each token through only part of its 2.4-trillion-parameter expert pool. Alibaba says 95 billion parameters are active for a token. That can reduce inference compute compared with a dense 2.4T model, but it does not turn the checkpoint into a 95B download. Unless a future runtime streams experts from storage or distributes them across machines, the complete expert set still has to be available.

This distinction matters for every local-hardware estimate. Treating Qwen3.8-Max as a normal 95B model would understate raw weight storage by roughly 25 times.

Preview versus production release

The qwen3.8-max-preview endpoint appeared on July 19 through Alibaba’s token plans and Qoder.¹ Alibaba’s documentation warned that the preview could be updated, removed, or replaced. The production qwen3.8-max launch followed on August 3, 2026 (general availability; announced August 2).¹²

At the August 3 snapshot, there was no final weight repository, download manifest, model card, or license to verify. “Open weights announced” was therefore accurate for that historical date; the official 2.4T³ and FP8⁴ repositories are public now under the custom Qwen3.8-Max license.⁵

Context, modalities, and API features

QwenCloud lists text, image, and video input with text output.⁶ Its headline context window is 1 million tokens, with more specific limits of 991K standard input, 983K thinking-mode input, 131K output, and up to 262K reasoning tokens.⁷ These caps apply separately, not additively.

The hosted service advertises function calling, structured outputs, context caching, and web search,⁸⁹ Built-in Responses API tools include a code interpreter, web extractor, web search, and image-search tools. Published service limits are 2 million tokens per minute and 15,000 requests per minute.⁶

Alibaba documents OpenAI-compatible, Anthropic-compatible, and DashScope access across multiple regions, including Beijing, Singapore, Tokyo, Frankfurt, and Virginia.¹⁰ Enterprises should still validate the exact regional endpoint, retention terms, data residency, and contractual controls before sending sensitive data.

Qwen3.8-Max API pricing

Current regional note (August 25, 2026): The [current Alibaba Model Studio page]¹¹ lists $1.65 input / $4.951 output per 1M tokens for Frankfurt (Germany), the US, Tokyo (Japan), Hong Kong, and Beijing; the Singapore endpoint and the current [QwenCloud model page]⁶ still show $2/$6. The table below therefore remains explicitly a QwenCloud/international snapshot from August 3. Check region, cache tier, taxes, and contract terms before production use.

QwenCloud’s list prices on August 3 are:⁶

Billing unit	USD per 1M tokens
Standard input	$2.00
Output	$6.00
Implicit cached input	$0.25
Explicit cache creation	$2.50
Explicit cache read	$0.17

Source: QwenCloud. Taxes and third-party provider markups are not included.

The calculation is:

cost = input_M×2 + output_M×6 + implicit_cache_M×0.25 + cache_create_M×2.50 + cache_read_M×0.17

Example workloads

Workload	Assumption	Calculated cost
small	1M input + 0.2M output	$3.20
medium	20M input + 5M output	$70
heavy	200M input + 50M output	$700

A repeated 100M-token prefix billed entirely as implicit cache hits would cost $25 instead of $200 in standard input charges, a theoretical $175 difference. Real savings depend on cache eligibility and prefix matching.

Can the full model run locally on a Mac?

Raw weight storage is parameter count multiplied by bits per parameter, divided by eight:

Weight format	Raw size for 2.4T parameters	Fits in 512 GB?
16-bit	4.8 TB	no
8-bit	2.4 TB	no
4-bit	1.2 TB	no
2-bit	600 GB	no

These figures exclude quantization metadata, runtime buffers, routing structures, and KV cache. Apple’s M3 Ultra Mac Studio tops out at 512 GB of unified memory.¹²¹³ The full Qwen3.8-Max checkpoint therefore does not fit on a single current Mac even under the unrealistic assumption of clean 2-bit weights with zero overhead.

The model could eventually be experimented with through multi-node sharding, SSD offload, or expert streaming. Those are specialized distributed systems, not the responsive local-LLM workflow most Mac users expect. Qwen3.8-27B remains the relevant local target for Mac testing.¹⁴ The full 2.4T checkpoint is now public, but it is still too large for a single Mac even under the theoretical 2-bit calculation; repository availability is not a local feasibility guarantee. First conversions of the 27B model exist as FP8,¹⁵ MLX 4-bit¹⁶ and an Unsloth GGUF;¹⁷ actual runtime behavior on a specific Mac still needs measurement.

What early benchmarks do—and do not—show

On Arena’s Software & IT Services text leaderboard dated August 1, Qwen3.8-Max was provisionally ranked ninth with a 1520±16 score from 1,405 votes.¹⁸ Its listed rank spread was 1–43. That is a promising early signal, but the uncertainty is large and the vote count is much smaller than for established models.

Alibaba also reports strong coding, multimodal, document, and long-horizon agent results. Those should remain labeled as vendor benchmarks until prompt sets, harnesses, model settings, and independent replications are available. An internal benchmark such as RecreationBench is not directly comparable with a widely reproduced public suite.

A separate Trilogy AI test ran the preview model and Kimi K3 against a matched 269-file software-architecture task.¹⁹ The models converged on the same core integration design and finished three blind-review points apart. Qwen was judged stronger on system boundaries; Kimi’s route was faster and more token-efficient. This is useful practitioner evidence, but it is one task against a mutable preview endpoint—not a general ranking.

Is qwen/qwen3.8-max an OpenRouter model ID?

Yes. The slug qwen/qwen3.8-max has been listed in OpenRouter’s public catalog since August 4, 2026 (released August 3, 2026).²⁰ OpenRouter forwards requests to Alibaba Cloud International at the same $2/$6 per-1M-token list price with a 1M context window, making it a confirmed, additional hosted access path alongside QwenCloud and Alibaba Model Studio.

OpenAI-compatible Python example

import os
from openai import OpenAI

key = os.getenv("DASHSCOPE_API_KEY")
if not key:
    raise RuntimeError("DASHSCOPE_API_KEY is not set")

client = OpenAI(
    api_key=key,
    base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
)

result = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[
        {"role": "user", "content": "Audit this design and identify the three highest-risk assumptions."}
    ],
)

print(result.choices[0].message.content)

Confirm the endpoint for the intended Alibaba region and API product before production use.

Practical status for Mac users

Qwen3.8-Max is now relevant both as a hosted model and as a public 2.4T checkpoint repository. The hosted version combines sparse 95B-per-token activation, multimodal input, a 1M context window, and the documented pricing; the open checkpoint is text-only, thinking-only, and still too large for a single Mac. For local evaluation, Qwen3.8-27B remains the relevant model.

For Mac users, the practical path is cloud access for Max and local evaluation of Qwen3.8-27B, whose official checkpoint and first conversions are now available. A reliable local assessment still requires independent long-context and agent benchmarks.

Sources

Footnotes

1. Qwen launch announcement for Qwen3.8-Max ↩ ↩²

2. Reuters on Alibaba’s announcement ↩

3. HF: Qwen3.8-2.4T-A95B ↩

4. HF: Qwen3.8-2.4T-A95B-FP8 ↩

5. Qwen3.8-Max license text ↩

6. QwenCloud model page (pricing, limits) ↩ ↩² ↩³ ↩⁴

7. Alibaba Model Studio: deep thinking ↩

8. Alibaba Model Studio: context cache ↩

9. Alibaba Model Studio: pricing ↩

10. Alibaba Model Studio: models ↩

11. Alibaba Model Studio: Qwen3.8-Max (regional pricing and limits) ↩

12. Apple Newsroom: Mac Studio ↩

13. Apple: Mac Studio tech specs ↩

14. HF: Qwen3.8-27B ↩

15. HF: Qwen3.8-27B-FP8 ↩

16. MLX community: Qwen3.8-27B-4bit ↩

17. Unsloth: Qwen3.8-27B GGUF ↩

18. Arena leaderboard, Software & IT Services ↩

19. Trilogy AI independent hands-on test ↩

20. OpenRouter catalog ↩