Qwen3.8-2.4T-A95B
Qwen3.8-2.4T-A95B is the open-weight release of Alibaba's Qwen3.8-Max, published in August 2026. It is a 512-expert mixture-of-experts with about 2.4T total and 95B active parameters, using a hybrid of Gated DeltaNet linear attention and gated full attention. The open checkpoint is text-only and always reasons before answering.
Specifications
- Context window
- 256Kabout 197,000 words, or 390 pages
- Max output
- Not published
- Input price
- $2per 1M tokens
- Output price
- $6per 1M tokens
- Cached input
- $0.25per 1M tokens
- Weights
- OpenQwen3.8-Max License
- Parameters
- 2.45T total, 95B activeMoE
- Input types
- Text
- Output types
- Text
- Reasoning mode
- Yes
- Tool calling
- Yes
- Released
- Aug 12, 2026Current
Together AI serverless price for Qwen/Qwen3.8-2.4T-A95B ($0.25 cached input per the Together pricing page). Fireworks lists the same $2.00 / $0.25 / $6.00 under the name Qwen 3.8 Max; DeepInfra lists Qwen3.8-Max at $1.65 / $4.951. Context is 262,144 native, extensible to about 1M with YaRN.
What it costs in practice
| Workload | Monthly cost at list price |
|---|---|
| Support chatbot30,000 conversations a month, 2,000 input and 400 output tokens each | $192 |
| Document summaries2,000 documents a month, 12,000 input and 800 output tokens each | $58 |
| Coding agent1,000 tasks a month, 60,000 input and 6,000 output tokens each | $156 |
Standard rates, no caching or batch discounts. Run your own numbers.
Strengths
- Max-class Qwen model with downloadable weights
- Hybrid linear attention keeps long-context KV cache small
- 262K native context, extensible to about 1M
Trade-offs to weigh
- Model-as-a-Service and AI-assistant businesses above $50M annual revenue need a separate license
- 2.4T parameters require a large multi-node deployment
Where it fits
Alibaba (Qwen) positions Qwen3.8-2.4T-A95B for:
Where you can use it
- Hugging Face
- Together AI
- Fireworks
- DeepInfra
Run it yourself
Released under Qwen3.8-Max License. Approximate memory needed just to load the weights:
| Precision | Memory for weights |
|---|---|
| 16-bit (BF16/FP16) | 4,892 GB |
| 8-bit (FP8/INT8) | 2,446 GB |
| 4-bit (Q4_K_M, about 4.85 bits) | 1,483 GB |
This is a mixture-of-experts model: all 2.45T total parameters must sit in memory, but only about 95B are used per token, so it runs faster than a dense model of the same size.
Add memory for the KV cache (it grows with context length and concurrent users) and 1-2 GB of runtime overhead. Size it precisely in the VRAM calculator.
Alternatives to consider
Frequently asked questions
How much does Qwen3.8-2.4T-A95B cost?
Qwen3.8-2.4T-A95B lists at $2 per million input tokens and $6 per million output tokens, with cached input at $0.25 (checked October 10, 2026). Together AI serverless price for Qwen/Qwen3.8-2.4T-A95B ($0.25 cached input per the Together pricing page). Fireworks lists the same $2.00 / $0.25 / $6.00 under the name Qwen 3.8 Max; DeepInfra lists Qwen3.8-Max at $1.65 / $4.951. Context is 262,144 native, extensible to about 1M with YaRN. Confirm current rates on the provider's pricing page before you commit.
What is the context window of Qwen3.8-2.4T-A95B?
Qwen3.8-2.4T-A95B accepts up to 262,144 tokens of context (about 197,000 words, or 390 pages). Long prompts cost more and can be slower, so most apps send only what the task needs.
Is Qwen3.8-2.4T-A95B open source?
Qwen3.8-2.4T-A95B is an open-weight model released under the Qwen3.8-Max License license. You can download the weights from Hugging Face (Qwen/Qwen3.8-2.4T-A95B) and run or fine-tune them yourself, within the license terms. Open weights isn't the same as fully open source: training data and code may not be published.
Can I run Qwen3.8-2.4T-A95B locally?
Yes, if you have enough memory. The weights alone need about 1,483 GB at 4-bit quantization and about 4,892 GB at 16-bit, plus room for the KV cache. Use the VRAM calculator on LLModel.com to check a specific GPU or Mac.
What is Qwen3.8-2.4T-A95B best for?
Alibaba (Qwen) positions Qwen3.8-2.4T-A95B for reasoning, coding, agents, research. It accepts text as input, has a reasoning (thinking) mode and supports tool calling. Test it on a sample of your own tasks before you choose it for production.
Sources
- huggingface.co/Qwen/Qwen3.8-2.4T-A95B
- docs.together.ai/docs/serverless-models
- github.com/QwenLM/Qwen3.8
- huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/config.json
- huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/LICENSE
- huggingface.co/Qwen/Qwen3.8-2.4T-A95B/raw/main/chat_template.jinja
- huggingface.co/api/models/Qwen/Qwen3.8-2.4T-A95B
- www.together.ai/pricing
- docs.fireworks.ai/serverless/pricing
- deepinfra.com/pricing
Facts checked against these pages on October 10, 2026. Spot an error? Tell us.
Still choosing a model?
Send us your use case and constraints. We'll come back with a shortlist and an evaluation plan you can run on your own data.
- Free and vendor-neutral: no model provider pays for placement.
- Your details go only to partners you agree to.