Downloads · 30 days
25
20% of all-time downloads
neopolita/Qwen3.6-35B-A3B-Saikei-8bit-mlx
Qwen3.6-35B-A3B-Saikei-8bit-mlx is a text generation model from neopolita. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as apache-2.0.
Downloads · 30 days
25
20% of all-time downloads
All-time downloads
125
Public
Parameters
35.1B
37.7 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors37.7 GB · 100%
How the weights are stored.
U3234.7B · 99%
From the Hugging Face model README

Qwen3.6-35B-A3B, un-pruned and improved: the intact 8-bit model after a
short knowledge-distillation pass from the 2.4-trillion-parameter
Qwen3.8-2.4T-A95B. Same size, same architecture, same speed — better
model. Stock mlx_lm, no custom code.
Saikei (栽景) are Japan's planted landscapes — miniature living scenes grown whole on a tray, where nothing is cut away: the art is in the cultivation. Where our Niwaki releases prune the tree, this model prunes nothing. It is the reference model, refined in place.
A paper with the full method and measurements is coming soon.
The unmodified 35B/3B-active MoE distilled against top-20 teacher logprobs from Qwen3.8-2.4T-A95B (tokenizer-identical): a 3M-token plan stopped at ~2.1M tokens when the validation probe plateaued; the released checkpoint is step 1769 (~1.8M tokens). Trainable during distillation: one linear output map per layer and the shared experts — 0.29B parameters, ~1.5 GPU-hours on one M4 Max. The maps were then folded back into the checkpoint, so the artifact is byte-compatible with the reference: 40 layers, 256 routed experts (top-8) + 1 shared per layer, everything 8-bit (group 64).
Screen protocol: WikiText-2 / C4 perplexity over 64 × 2048-token windows, paired same-day against the reference; 500-example/task suite (arc_easy, hellaswag, piqa, winogrande, boolq); gsm8k (100 problems, lm-eval's default few-shot prompt, strict extraction; the chat-template/think-mode run we intended did not take effect — a runner flag error found in review).
| model | wt2 ppl ↓ | c4 ppl ↓ | task avg ↑ | gsm8k ↑ |
|---|---|---|---|---|
| reference (8-bit) | 6.61 | 10.62 | 0.790 | 0.99 |
| this model | 6.33 | 11.30 | 0.817 | 0.99 |
All four numbers were measured on this released artifact (perplexities on 2026-08-27; the task suite and gsm8k on 2026-09-06 through a rebuild whose shards are SHA-256-identical to the published files). The mid-run checkpoint 1182 of the same heal, evaluated during development, scored 6.32 / 11.18 / 0.814.
The trade-off, stated honestly: WikiText-2 and the task suite improve; C4 drifts up toward the teacher's distillation-corpus mix; reasoning (gsm8k, base-prompt protocol) is exactly preserved. If your use leans on broad web-crawl-style text, the reference may serve you equally well; for knowledge-and-reasoning workloads the distilled model measures better across the board.
| total / active parameters | 35B / ~3.8B (unchanged) |
| layers / routed experts / top-k | 40 / 256 / 8 (unchanged) |
| precision | 8-bit throughout (group 64, unchanged) |
| whole-model size on disk | ~37.7 GB |
| runs comfortably in | 48 GB+ unified memory |
pip install mlx-lm
mlx_lm.generate --model neopolita/Qwen3.6-35B-A3B-Saikei-8bit-mlx \
--prompt "Explain speculative decoding in two paragraphs." -m 600
Built from mlx-community/Qwen3.6-35B-A3B-8bit (untouched backbone,
routed experts, router, embeddings) + a distillation heal of the shared
experts and per-layer output maps, teacher Qwen/Qwen3.8-2.4T-A95B.
The saikei block in config.json records the recipe.