Hugging FaceBuilt to carry complex, multi-step tasks through to the end, reasoning before every answer.
- Rating
- 4.3
- Downloads
- 6.7K
- Parameters
- 2.4T, 95B active
- Context
- 256K
- Download
- 397 to 4893 GB
- Commercial use
- With conditions
Too big for this Mac
Its smallest size needs 736 GB of memory, more than M4 · 16 GB can give it.
About
Qwen says it improved planning on its own and handling feedback during agent tasks. It is built on the design of the earlier Qwen3.5 models.
Limitations
It reads text only and can't take pictures or other input.
Its reasoning step always runs and can't be switched off.
Good to know
Qwen says this is the first time it has openly released a model of its Max class.
Sources: Qwen/Qwen3.8-2.4T-A95B model card. Every line comes from one of these.
Good at
Licence
Qwen3.8-Max License. Commercial use: with conditions. MIT-style license. A product or service with more than 100 million monthly active users or $20 million in monthly revenue must display the model name on its interface. A Model-as-a-Service or AI Work Assistant business whose group revenue passes $50 million in any 12 months needs a separate license from Qwen before commercial use. The maker's page has the full terms.
Qwen3.8 2.4T runs well from a Mac with an M3 Ultra and 512 GB of memory, at UD-Q1_0, writing about 13 words (17 tokens) a second. 2 of 59 Mac configurations run it well.
Compare with
The nearest models in size, judged for the same Mac.
- Kimi K3Long coding work and picturesToo big for this Mac · 712 GBGet
- DeepSeek V4 ProBuilt for long, many-step tasks that need a lot of text held in view at once.Too big for this Mac · 850 GBGet
- Kimi K2.6Long coding jobs and agentsToo big for this Mac · 341 GBGet
- GLM 5.3Whole software projectsToo big for this Mac · 254 GBGet
Sizes
| Size | Download | Memory at 8K | Memory at 32K | Open |
|---|---|---|---|---|
| UD-IQ2_XSSmaller | 731 GB | 722 GB | 723 GB | Open |
| UD-IQ3_XXSBalanced | 956 GB | 944 GB | 946 GB | Open |
| UD-IQ4_XSBest | 1311 GB | 1296 GB | 1298 GB | Open |
All 9 sizes
Memory is what llama.cpp's own calculator reported for each size when the catalog was built, with the app's settings. Smaller, Balanced and Best are the three sizes the store names.
Run It Anywhere
Qwen3.8 2.4T is an open model, so any app that runs GGUF files can run it. These take the UD-IQ3_XXS size from unsloth straight from Hugging Face.
- LM StudioMac, Windows, Linux
Open - JanMac, Windows, Linux
Open - MstyMac, Windows, Linux
Open - OllamaMac, Windows, Linux
ollama run hf.co/unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ3_XXS - llama.cppMac, Windows, Linux
llama-server -hf unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ3_XXS - Docker Model RunnerMac, Windows, Linux
docker model run hf.co/unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ3_XXS
On a PC Graphics Card
| Video memory | Largest size that fits | Download | Memory at 8K |
|---|---|---|---|
| 6 GB | None worth running | needs 944 GB | |
| 8 GB | None worth running | needs 944 GB | |
| 10 GB | None worth running | needs 944 GB | |
| 11 GB | None worth running | needs 944 GB | |
| 12 GB | None worth running | needs 944 GB | |
| 16 GB | None worth running | needs 944 GB | |
| 20 GB | None worth running | needs 944 GB | |
| 24 GB | None worth running | needs 944 GB | |
| 32 GB | None worth running | needs 944 GB | |
| 48 GB | None worth running | needs 944 GB |
The largest size, of those that keep three bits a weight or more, whose memory at an 8K chat is under 92% of the card's; the rest is the desktop's. Estimated from the files, the same way as for a Mac; speed on a PC is not estimated yet. A size that doesn't fit the card still runs, slower, with part of it in ordinary memory.
Tests
GPQA Diamond: questions PhD scientists write for each other. Humanity's Last Exam: questions written to be beyond today's models. Scores are the ones filed on the model's Hugging Face page or the board's own table, read 2026-11-05; some come from the maker.
On every Mac
| Mac | Size | Fit | Memory | Writes | Reads |
|---|---|---|---|---|---|
| M6 · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M6 · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M6 · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Pro · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Pro · 48 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Pro · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Max · 36 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Max · 48 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Max · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Max · 128 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Ultra · 96 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Ultra · 256 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M5 Ultra · 512 GB | UD-Q1_0 | Runs well | 402 GB | 23 | 171 |
| M4 · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Pro · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Pro · 48 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Pro · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Max · 36 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Max · 48 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Max · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Max · 96 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M4 Max · 128 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 · 8 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Pro · 18 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Pro · 36 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Max · 36 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Max · 48 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Max · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Max · 96 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Max · 128 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Ultra · 96 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Ultra · 256 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M3 Ultra · 512 GB | UD-Q1_0 | Runs well | 402 GB | 13 | 48 |
| M2 · 8 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 · 24 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Pro · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Pro · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Max · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Max · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Max · 96 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Ultra · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Ultra · 128 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M2 Ultra · 192 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 · 8 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 Pro · 16 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 Pro · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 Max · 32 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 Max · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 Ultra · 64 GB | UD-IQ2_XS | Too big | 736 GB | ||
| M1 Ultra · 128 GB | UD-IQ2_XS | Too big | 736 GB |
Writes and reads are words a second (a token is about three quarters of a word), 4K tokens into a chat, estimated unless marked measured. The size is the one the store offers that Mac. Constants fitted October 11, 2026; how, and how wrong.
Badge for the model card
Shows on Qwen3.8 2.4T's Hugging Face card which Macs run it well, linked to this page. It updates every morning.
Sizes and files are read from unsloth/Qwen3.8-2.4T-A95B-GGUF on Hugging Face and checked every morning. LM Arena has not rated this model, so the rating is Outmake's estimate from its size and release date; how the estimate is made. Catalog of November 4, 2026.