Qwen3-Next-80B-A3B-Instruct for Janas-LLM
On a laptop running Debian 13 Linux, with an Intel Core Ultra 9 185H, its integrated Arc GPU and 32 GB of memory, Janas-LLM reads a prompt at 62 tok/s and writes at 24 tok/s, 37 tok/s with its own draft block in a chat: llama.cpp, on the same laptop, 16 and 2.8.
| On the same laptop |
Janas-LLM |
llama.cpp |
| reading a prompt of 1,024 tokens |
62 tok/s |
16 tok/s |
| writing (16 tokens) |
24 tok/s |
2.8 tok/s |
| writing in a chat, with its own draft block |
37 tok/s |
not measured |
Measured on 9-11 October 2026: llama-bench -p 1024 -n 16 and janas-bench --prompt 1024 --gen 16, both at their best configuration; the drafts with janas-bench on a chat reply, the mean of three rounds (a draft guesses the next tokens and the model checks them: the same text, faster). On other machines the figures will differ. Details: Janas README.
The machine these figures come from
- Debian 13 (trixie) Linux, x86-64, kernel 6.12.107
- Intel Core Ultra 9 185H: 6 performance cores, 8 efficiency cores, 2 low-power cores; AVX2, FMA, F16C, AVX-VNNI
- its integrated Intel Arc GPU, through the Vulkan driver of Mesa 26.1.6
- 32 GB of DDR5-5600 memory
- a 1 TB NVMe SSD, about 5.7 GB/s reading
- llama.cpp build
2b18470 (18 September 2026), Vulkan backend
The model (48.4 GB) is larger than the laptop's memory: Janas-LLM streams from the NVMe SSD the experts each token needs and keeps the ones that keep coming back in memory. llama.cpp with its usual build ran out of memory on it; the figures above are from a build without weight repacking.
These are the model's weights converted to JNS, the format of Janas-LLM, the inference engine in C of Janas, for ordinary computers, with or without a GPU. The files are read by Janas-LLM only.
Get it
Janas-LLM for Linux x86-64 (Ubuntu 22.04, Debian 12 and later), then the model:
curl -LO https://github.com/prabanta-dev/janas/releases/latest/download/janas-linux-x86_64.tar.gz
tar xf janas-linux-x86_64.tar.gz && cd janas
./janas-get qwen3-next-80b-a3b
./janas-chat qwen3-next-80b-a3b
janas-get downloads the files of this repository and checks them against the fingerprints below; janas-chat with no model shows the catalog and downloads the one you choose. Building Janas-LLM from its sources instead: prabanta-dev/janas.
Files
| File |
Size |
SHA-256 |
Converted from |
qwen3-next-80b-a3b-q4km.jns |
48.41 GB |
fa3255c73106391e9cb5a97efd30f77a928fb4518d40060f795a8cff6656daed |
Qwen/Qwen3-Next-80B-A3B-Instruct-GGUF Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf (catalog name qwen3-next-80b-a3b: Qwen3-Next-80B-A3B-Instruct, with its MTP block) |
qwen3-next-80b-a3b-mtp-q4k.jns |
0.94 GB |
5c9bc4cb6f739193fac5e4daa30c922f73255acf41443ecfc0123ef084584559 |
the multi-token prediction tensors of Qwen/Qwen3-Next-80B-A3B-Instruct |
What was changed
These files are a modified work of the original model, as the Apache License 2.0 requires to be stated. Nothing was retrained or pruned: each file is a GGUF file quantized by the people named above, rewritten by Janas-LLM's gguf2jns in the order Janas reads it, with the metadata carried over. Each (layer, expert) has a slot of its own, so that the experts a token needs can be read from disk on their own. The experts' down matrices were then cut by Janas-LLM's jns_planes into three planes of two bits each: a reordering of the same bits, which with all three planes gives back the quantized weights to the last bit. The multi-token prediction block, not in the GGUF, was taken from the original checkpoint and quantized to Q4_K by Janas-LLM's hf2jns_mtp.
Licence and attribution
Licensed under the Apache License 2.0, as the original model: see LICENSE.
LICENSE is the original repository's own.
