Skip to page content
SAVRN
Search Plan a project

Open-weight model

Qwen3-Coder-Next-moved-store

by Jalen Lu jalenluorion/Qwen3-Coder-Next-moved-store

Qwen3-Coder-Next-moved-store is an open-weight model from Jalen Lu, released under Apache License 2.0. It has 262,144-token context. Its published files total 84.3 GB.

Qwen/Qwen3-Coder-Next (Apache-2.0), the agentic-coding model on the Qwen3-Next 80B-A3B architecture, re-laid-out so that only 3% of the parameters have to be in memory.

Parameters—
Context262,144
Weights84.3 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Model Card

By Jalen Lu, published under apache-2.0, revision 6eb0ed7c7373.

Qwen/Qwen3-Coder-Next (Apache-2.0), the agentic-coding model on the Qwen3-Next 80B-A3B architecture, re-laid-out so that only 3% of the parameters have to be in memory. Every routed expert is stored exactly once on disk as int8 with one scale per neuron; everything else stays at the released precision; nothing was retrained. The multi-token-prediction head of the release is not included. The layout, runner and mechanism are those of whose card explains them and carries the measurements: small routers read the current token's state four layers early and name the experts each layer will want, so the reads overlap the computation of the layers in between; the released router then picks its…

Read Jalen Lu's full model card

Qwen3-Coder-Next, experts moved to disk

Qwen/Qwen3-Coder-Next (Apache-2.0), the agentic-coding model on the Qwen3-Next 80B-A3B architecture, re-laid-out so that only 3% of the parameters have to be in memory. Every routed expert is stored exactly once on disk as int8 with one scale per neuron; everything else stays at the released precision; nothing was retrained. The multi-token-prediction head of the release is not included.

part parameters where
attention and linear-attention layers, shared experts, norms, embeddings, output head, the released router gates (resident.safetensors) 2.36B memory (3.8 GB in fp16 with the embedding table read from disk)
24,576 routed experts (store/layer_XX.bin, int8, one scale per neuron, 3.0 MiB records) 77.3B disk, 72.4 GiB

The layout, runner and mechanism are those of jalenluorion/Qwen3-Next-80B-A3B-moved-store, whose card explains them and carries the measurements: small routers read the current token's state four layers early and name the experts each layer will want, so the reads overlap the computation of the layers in between; the released router then picks its top-10 among the experts in memory. The routers for this model are being trained; until routers_k4.pt appears here, run the store in exact mode (--mode exact: the released routing, missing experts read on the spot), which is slow but exact.

pip install mlx mlx-lm tokenizers jinja2 numpy huggingface_hub
hf download jalenluorion/Qwen3-Coder-Next-moved-store --local-dir coder-next-store
cd coder-next-store
python mlx_run.py --mode exact                      # until the routers are published
python mlx_run.py                                   # prefetch mode, once routers_k4.pt is here
python mlx_run.py --serve 8080 --prompt-cache demo/cache --warm   # OpenAI-compatible API for OpenCode

sha256.txt holds checksums of the first store layer and the resident file; convert.py rebuilds everything from the released weights.

Configuration

Architecture
Qwen3NextForCausalLM
Context length (tokens)
262,144
Layers
48
Hidden size
2,048
Feed-forward size
5,120
Attention heads
16
Key/value heads
2
Head dimension
256
Vocabulary size
151,936
Experts
512
Experts active per token
10
RoPE base
5,000,000
Stored precision
bfloat16
Model type
qwen3_next

Identity and Version

Repository
jalenluorion/Qwen3-Coder-Next-moved-store
Publisher
Jalen Lu
Task
Not stated by the source
Modality
Other
Library
Not stated by the source
Parameters
Not stated by the source
Languages
Not stated by the source
Revision
6eb0ed7c73731a5eb65a6e0f2d8a01e956a88b9c
First published
2026-10-15
Last updated
2026-10-16

Files and Weights

64 files, 84.3 GB in total. The weights are 50 files totalling 84.3 GB in bin, pt, safetensors.

Weights50 files · 84.3 GB
Configuration6 files · 82.5 KB
Tokenizer4 files · 11.5 MB
Documentation1 file · 2.3 KB
Other2 files · 1.4 KB
Repository1 file · 1.5 KB
Every file
FileTypeSizeSHA-256
resident.safetensorsWeights4.7 GB 323903ad5a1f
routers3/coder_next_router3/ckpt.ptWeights1.9 GB c28361dfa8ca
store/layer_00.binWeights1.6 GB 38e3f8434ace
store/layer_01.binWeights1.6 GB 9da2867a4f93
store/layer_02.binWeights1.6 GB 2f16d512e3a5
store/layer_03.binWeights1.6 GB 65fae7957025
store/layer_04.binWeights1.6 GB 83a91156035b
store/layer_05.binWeights1.6 GB 525a83510a09
store/layer_06.binWeights1.6 GB fdd11e092e48
store/layer_07.binWeights1.6 GB c0af3ff47c5e
store/layer_08.binWeights1.6 GB af847b675060
store/layer_09.binWeights1.6 GB 4f4d8e75673c
store/layer_10.binWeights1.6 GB c87ff073498e
store/layer_11.binWeights1.6 GB 6ec5b67cd929
store/layer_12.binWeights1.6 GB 4432e8f9d70f
store/layer_13.binWeights1.6 GB a123f7fd8b69
store/layer_14.binWeights1.6 GB 85b13267c9b6
store/layer_15.binWeights1.6 GB cf60f8dd7ffe
store/layer_16.binWeights1.6 GB 57a73b3a64cf
store/layer_17.binWeights1.6 GB 38bd617228ed
store/layer_18.binWeights1.6 GB 466fdd83164e
store/layer_19.binWeights1.6 GB 0356c9b79cd6
store/layer_20.binWeights1.6 GB 091d047a891c
store/layer_21.binWeights1.6 GB 1ba7ce36a838
store/layer_22.binWeights1.6 GB c0e85f073895
store/layer_23.binWeights1.6 GB fa88195cba7c
store/layer_24.binWeights1.6 GB 71f40292c4ec
store/layer_25.binWeights1.6 GB 9a037dfbd29c
store/layer_26.binWeights1.6 GB 89f41c47e417
store/layer_27.binWeights1.6 GB dc076fd10e87
store/layer_28.binWeights1.6 GB 6008d055d47c
store/layer_29.binWeights1.6 GB 2df97482405e
store/layer_30.binWeights1.6 GB 0e475c5aa3ff
store/layer_31.binWeights1.6 GB 53c668bdd4ea
store/layer_32.binWeights1.6 GB f14fb15224ea
store/layer_33.binWeights1.6 GB f72e39771437
store/layer_34.binWeights1.6 GB ca04107bf08d
store/layer_35.binWeights1.6 GB 03ec4af1c1cf
store/layer_36.binWeights1.6 GB 87917ad54123
store/layer_37.binWeights1.6 GB e1b496bd769d
store/layer_38.binWeights1.6 GB f7a5519d7b63
store/layer_39.binWeights1.6 GB b38efb2f33df
store/layer_40.binWeights1.6 GB 8c8cc66542d3
store/layer_41.binWeights1.6 GB 0eb25390fe80
store/layer_42.binWeights1.6 GB d1fdf7d04fb2
store/layer_43.binWeights1.6 GB cdbf7f8a9e12
store/layer_44.binWeights1.6 GB 10b6eabd94d4
store/layer_45.binWeights1.6 GB cd13862cd01f
store/layer_46.binWeights1.6 GB d677421d0002
store/layer_47.binWeights1.6 GB 3428e8b6d7b9
config.jsonConfiguration1.2 KB —
convert.pyConfiguration10.2 KB —
demo/opencode.jsonConfiguration1.3 KB —
generation_config.jsonConfiguration214 B —
mlx_run.pyConfiguration69.1 KB —
store/meta.jsonConfiguration525 B —
README.mdDocumentation2.3 KB —
demo/demo.shOther1.2 KB —
sha256.txtOther172 B —
.gitattributesRepository1.5 KB —
merges.txtTokenizer1.7 MB —
tokenizer.jsonTokenizer7.0 MB —
tokenizer_config.jsonTokenizer11.7 KB —
vocab.jsonTokenizer2.8 MB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
84.3 GB
Download from Jalen Lu

Released by Jalen Lu through its official repository on Hugging Face. Read the license.

Built From

  • Derived from Qwen/Qwen3-Coder-Next

Memory Requirements

PrecisionWeights in memory
As published84.3 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About Qwen3-Coder-Next-moved-store

Can I use Qwen3-Coder-Next-moved-store commercially?

Yes. Qwen3-Coder-Next-moved-store is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

What is Qwen3-Coder-Next-moved-store's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.