Qwen4-Max: Run It Locally? Ollama, API and Pricing (2026)
No, you cannot run Qwen4-Max locally yet. Alibaba announced the 2.4-trillion-parameter model on November 16, 2026 as a hosted release only. As of November 17, 2026 there are no downloadable weights, no Ollama library page, and no Qwen4 repository on the official Qwen Hugging Face organization.
You can use it today through QwenCloud, QwenWork, Qoder, QwenStudio and OpenRouter. This guide covers what Alibaba says about the model, the four checks that show local use is off the table for now, the API setup that works today, what it costs, and a Qwen model you can run on Ollama while you wait for an open Qwen4.

Qwen4-Max at a Glance
Qwen4-Max is Alibaba's new flagship model, announced on the official Qwen blog on November 16, 2026. It is the first model in the Qwen4 family. Alibaba says smaller Qwen4 models are coming to open source later, but the flagship itself is hosted.
| Spec | Qwen4-Max |
|---|---|
| Announced | November 16, 2026 |
| Backbone parameters | 2.4 trillion |
| Active per token | 71 billion |
| Extra N-gram embedding parameters | 328 billion |
| Context window | 1 million tokens |
| Input | Text, image, video |
| Output | Text |
| API model ID | qwen4-max |
| Weights | Not released |
| Ollama library | Not listed |
| Price (International) | $1.50 in, $4.50 out per 1M tokens |
Sources: the official Qwen4-Max announcement and the Alibaba Cloud Model Studio model page. Prices are the International deployment rates listed on November 17, 2026.
The architecture is a sparse mixture-of-experts design with three changes Alibaba names. Gated DeltaNet layers paired with Qwen Sparse Attention cut the cost of long contexts. Four parallel gated residual branches improve training stability. An N-gram embedding layer adds capacity with little extra compute.
Alibaba says Qwen4-Max-Base reaches the lowest bits per byte of the five checkpoints it compared. It improves on Qwen3.8-Max-Base by about 5 to 13 percent while activating 25 percent fewer parameters.
Can You Run Qwen4-Max Locally? Four Checks Say No
Not today. Four separate places agree on that, all checked on November 17, 2026.
- •Ollama: the library pages at ollama.com/library/qwen4-max and ollama.com/library/qwen4 both return a 404.
- •Hugging Face: the official Qwen organization lists Qwen3.8 as its newest text models. No Qwen4 repository exists there.
- •Alibaba's docs: the Model Studio page for qwen4-max says batch inference, fine-tuning and model deployment are not supported. It describes the model only as a hosted API.
- •Alibaba's announcement: the blog says smaller models in the Qwen4 family are coming to open source. It names no size, license or date, and says nothing about releasing Max itself.
Watch for lookalike uploads. Third-party repositories with names like Qwen4-27B-GGUF and Qwen4-27B-Uncensored have already appeared on Hugging Face from accounts that are not Qwen's. Alibaba has not announced a Qwen4 27B model, so treat any file with that name as unverified. Download Qwen weights only from the official Qwen organization at huggingface.co/Qwen.
Why a local install would not work even with weights
The size alone rules out a normal machine. This is a back-of-the-envelope calculation, not an Alibaba figure, using the 2.4 trillion backbone parameters:
| Precision | Bytes per parameter | Memory for weights alone |
|---|---|---|
| 16-bit | 2 | about 4.8 TB |
| 8-bit | 1 | about 2.4 TB |
| 4-bit | 0.5 | about 1.2 TB |
That leaves out the extra 328 billion N-gram embedding parameters and the memory a long context needs. Even a 4-bit copy is far beyond any laptop, workstation or single-GPU server.
When open Qwen4 weights do arrive, the smaller models will be the realistic local option. Check the Ollama library for the exact model name and tag once a release is announced. Do not guess the tag.
More AI tool picks like this
Tested AI tools, honest verdicts, and real comparisons, delivered every Wednesday and Saturday.
What You Can Run on Ollama Today: Qwen3.8 27B
Qwen3.8 is the newest Qwen family on Ollama. The 27B build is an 18 GB download with a 256K context window and accepts text and images, according to the Ollama library page. Its tags list vision, tools and thinking.
ollama run qwen3.8
To point a coding agent at your local model, Ollama's library page shows the same launch pattern for Claude Code, OpenCode, Hermes Agent and OpenClaw:
ollama launch claude --model qwen3.8
New to Ollama? The Ollama setup guide covers the install, and the best local LLM models guide compares what fits which hardware. For what Reddit says about running models at home, see the local LLM guide.
Qwen3.8 27B is a much smaller model than Qwen4-Max. Do not expect Qwen4-Max results from it. The benchmark scores and the cost drop Alibaba reports belong to the flagship.
How to Use Qwen4-Max Today
Alibaba's announcement says Qwen4-Max is available on QwenCloud through either the Token Plan or the API. You can also use it on QwenWork, Qoder and QwenStudio. If you want access from your own code, there are three practical routes.
Option 1: OpenRouter, the fastest way to try it
OpenRouter lists the model as qwen/qwen4-max, served by Alibaba Cloud International, at $1.425 per million input tokens and $4.275 per million output tokens, per its model page. Its API reports a 1,000,000-token context. Set your OPENROUTER_API_KEY first.
curl -X POST https://openrouter.ai/api/v1/chat/completions -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model": "qwen/qwen4-max", "messages": [{"role": "user", "content": "Summarize this contract in five bullets."}]}'
Option 2: Alibaba Cloud Model Studio with the OpenAI SDK
Create a Model Studio API key, then point the OpenAI SDK at the base URL for the region your key belongs to. Keys are region-specific, and a mismatch returns a 401 error, per Alibaba's base URL reference.
| Region | OpenAI-compatible base URL |
|---|---|
| Singapore | https://dashscope-intl.aliyuncs.com/compatible-mode/v1 |
| US (Virginia) | https://dashscope-us.aliyuncs.com/compatible-mode/v1 |
| China (Beijing) | https://dashscope.aliyuncs.com/compatible-mode/v1 |
from openai import OpenAI client = OpenAI( api_key="YOUR_MODEL_STUDIO_KEY", base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1", ) response = client.chat.completions.create( model="qwen4-max", messages=[{"role": "user", "content": "Plan a three-day product launch checklist."}], reasoning_effort="high", ) print(response.choices[0].message.content)
Thinking mode is on by default, and reasoning_effort accepts low, high or xhigh, per the model page. The same page lists OpenAI Responses, Anthropic Messages and DashScope as supported protocols. Model Studio lists a free quota of 1 million tokens, valid for 90 days from activation.
Option 3: Claude Code through the Anthropic-compatible endpoint
Model Studio also exposes an Anthropic Messages endpoint, so Claude Code can send requests to a Qwen model. Alibaba's Claude Code guide uses qwen3.7-max as its example model. Swapping in qwen4-max should work because the model page lists the Anthropic Messages protocol, but that exact combination is not in Alibaba's guide. Test it with a short prompt first.
export ANTHROPIC_AUTH_TOKEN="YOUR_MODEL_STUDIO_KEY" export ANTHROPIC_BASE_URL="https://dashscope-intl.aliyuncs.com/apps/anthropic" export ANTHROPIC_MODEL="qwen4-max" claude -p "Hello"
Alibaba's guide also maps ANTHROPIC_DEFAULT_HAIKU_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL and CLAUDE_CODE_SUBAGENT_MODEL to Qwen models, so every model slot uses Qwen. Its main example uses a workspace-specific base URL, while its FAQ lists the dashscope-intl address above. Use the FAQ form if the workspace form returns an error.
Every route above sends your prompts to Alibaba Cloud servers, not to your machine. Do not send code or documents you are not allowed to share with a third party.
What Qwen4-Max Costs, and What the 62 Percent Claim Means
Qwen4-Max is priced per million tokens, and the rate depends on where you buy it.
| Route | Input per 1M | Output per 1M |
|---|---|---|
| Model Studio, International | $1.50 | $4.50 |
| Model Studio, Global | $1.413 | $4.24 |
| OpenRouter (Alibaba Cloud Int.) | $1.425 | $4.275 |
Cached input is cheaper. OpenRouter lists $0.07125 per million cached-read tokens.
The number most coverage skips
A prompt that fills the whole 1,000,000-token context window costs $1.50 in input tokens at the International rate. A 10,000-token answer adds $0.045. One full-context request with a long answer therefore costs about $1.55, and the free 1 million token quota covers the input of exactly one such prompt. This is our own arithmetic from the listed prices.
Alibaba's 62 percent claim
Alibaba says that on its own Work360 benchmark, the average cost per task fell from $4.43 to $1.67, a 62 percent drop compared with Qwen3.8-Max. It credits two causes: lower API prices, with large cuts to cached-input pricing, and fewer tokens used per task. Token use dropped about 24 percent at Xhigh effort and about 35 percent at High effort on Work360.
That is a vendor benchmark. No independent test of the figure was available when this guide was written. Measure cost per finished task on your own work, not price per token.
What Alibaba Says Qwen4-Max Does Better
These are Alibaba's claims from its announcement, not independent results.
- •Work360: Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max. Alibaba built the benchmark around a taxonomy of 26 top-level domains, 182 subdomains and 721 specialized fields.
- •Fewer tokens: on Terminal-Bench 4.0, token use drops about 48 percent at Xhigh effort and about 49 percent at High effort compared with Qwen3.8-Max.
- •Manufacturing demo: Alibaba says the manual effort to program one workpiece fell from about 475 minutes to 24 minutes, about 95 percent, in production environments at SANF and Pacific Union.
- •Legal demo: given 198 inconsistently named working papers, the model sorted them into 14 folders and produced an 86-page due diligence report with citations to the source papers.
- •Live Delivery: the agent hands over a usable first version early, then improves it in successive releases while the user steers. Alibaba is open-sourcing a Qwen-Live-Delivery-Plugin for it.
Treat the demos as vendor claims until independent testing appears.
Qwen4-Max vs Qwen3.8 27B on Ollama
The choice is between a hosted flagship and a model that stays on your machine.
| Qwen4-Max | Qwen3.8 27B | |
|---|---|---|
| Where it runs | Alibaba Cloud | Your machine |
| Context window | 1M tokens | 256K tokens |
| Download size | None, hosted | about 18 GB |
| Input | Text, image, video | Text, image |
| Cost | $1.50 in, $4.50 out per 1M | Free after hardware |
| Prompts leave your machine | Yes | No |
| Ollama command | Not available | ollama run qwen3.8 |
Pick Qwen4-Max when you need the 1 million token window or the strongest agent results and can send data to a cloud provider. Pick Qwen3.8 27B when privacy, offline use or zero per-token cost matters more.
When Will an Open Qwen4 Arrive?
Alibaba has not said. The announcement promises smaller open-source models in the Qwen4 family, with no sizes and no dates.
For context, the official Qwen Hugging Face listing includes a repository named Qwen3.8-2.4T-A95B alongside Qwen3.8-27B and Qwen3.8-Flash-Next. Even a model that size needs server-class memory, so the small models are the ones that matter for local use.
Three places will confirm a release first:
- •The official Qwen organization on Hugging Face, huggingface.co/Qwen
- •The Ollama library at ollama.com/library
- •The Qwen blog at qwen.ai/blog
This guide will be updated when a Qwen4 model appears on Ollama with a real tag and size.
Frequently Asked Questions
No. Alibaba released Qwen4-Max as a hosted model with no downloadable weights. The announcement says smaller models in the Qwen4 family are coming to open source, but it gives no sizes, license or dates, and it does not promise to release Max itself.
Qwen4-Max Is Cloud-Only for Now
Use Qwen4-Max through QwenCloud, OpenRouter or Model Studio, and run Qwen3.8 27B on Ollama if you need something local. The open Qwen4 models Alibaba promised have no date yet, and the first place they will appear is the official Qwen Hugging Face organization.
About the Author

Amara
Amara is an AI tools expert who has tested over 1,800 AI tools since 2022. She specializes in helping businesses and individuals discover the right AI solutions for text generation, image creation, video production, and automation. Her reviews are based on hands-on testing and real-world use cases, ensuring honest and practical recommendations.
View full author bio→Related Guides
Local LLM Reddit: What the Privacy-First AI Community Really Thinks in 2026
What does Reddit say about running local LLMs? Community verdict on Ollama vs LM Studio, best local models (Llama, Mistral, DeepSeek), hardware requirements, and privacy benefits.
DeepSeek vs ChatGPT Reddit: Chinese AI Disrupts OpenAI in 2026 (Community Verdict)
Compare DeepSeek and ChatGPT based on real community feedback. DeepSeek offers open-source efficiency and coding precision, ChatGPT provides creative polish and multimodal features.