Downloads · 30 days
6
22% of all-time downloads
qwen3.5-35b-topiary-stream
jesusluque/qwen3.5-35b-topiary-stream
qwen3.5-35b-topiary-stream is a machine learning model from jesusluque. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
A servable residency artifact for a model that does not fit whole in 24 GB: the 4-bit checkpoint loads at 19.5 GB and Metal fails on generation. Every expert is split into standard-kernel-readable bit planes (P0 floor…
All-time downloads
27
Public
Parameters
34.7B
19.5 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors11.5 GB · 59%
Parameter types
How the weights are stored.
U3234.7B · 100%
Base models
Model card
From the Hugging Face model README
qwen3.5-35b-topiary-stream — Topiary Stream artifact (resident-p0)
A servable residency artifact for a model that does not fit whole in 24 GB: the 4-bit checkpoint loads at 19.5 GB and Metal fails on generation. Every expert is split into standard-kernel-readable bit planes (P0 floor resident in the checkpoint, P1 completion in row-pageable memmaps), served by the Topiary Stream runtime with a gate-governed pool, a guaranteed quality floor and an optional elastic memory governor.
Serving caveat, stated plainly: this artifact does NOT load with stock
mlx_lm. It requires the Stream runtime (pure Python over stock MLX kernels
— no custom Metal):
python src/serve.py --artifact jesusluque/qwen3.5-35b-topiary-stream \
--pool-k 32 --governor
Measured results (Apple M5 Pro, 24 GB)
| Metric | Value |
|---|---|
| 4-bit size / served peak | 19.5 GB (unservable) / 12.5–13.9 GB |
| Decode speed | 44–47 tok/s sustained (K=64–32) |
| HumanEval / GSM8K (greedy) | 92% (23/25) / 92% (46/50) |
| With exact prefill (current runtime, n=15) | 14/15 / 15/15 · TF PPL 2.3614/7.0583 = base |
| Exact-pager control (τ=0) | PPL 2.3614 vs true base 2.3623 — no measurable difference |
| Operating point τ=0.10 | +0.9% code / +0.6% wiki PPL |
| WikiText PPL vs best natively-fitting model | 7.11–7.83 vs 10.27 (−24 to −31%) |
| MATH-500 / MBPP (n=100) · MMLU / LAMBADA (n=500) | 60% / 78% · 83.0% / 75.4% |
| Citable throughput (fresh reboot, 3 interleaved 1024-token rounds, swap-free) | 47.2 tok/s median (cold round within 1% of warm) |
Honest limits
The runtime serves prefill exact by design (the prompt's P1 union reads once from the memmaps), so the pool policy governs decode only; before that fix the pool's teacher-forced prefill toll measured +6–11% PPL (flat prefill routing defeats recency). At n=25/50, task differences of ≤2 items are indistinguishable under exact McNemar; treat the 92/92 as indistinguishable from the best fitting model, not proven equal. Throughput is now measured from a clean reboot (47.2 tok/s, swap-free).
Provenance
Split deterministically from the community 4-bit conversion of
Qwen/Qwen3.5-35B-A3B with split.py --layout resident-p0; plane arithmetic
verified exact (2.6e-7, float accumulation order) against stock kernels in the repo's CI. Full measurement
history including negative results:
https://github.com/jesusluque/topiary-stream
- Library
- mlx
- Type
- qwen3_5_moe
- License
- apache-2.0
- Created
- Aug 16, 2026
- Updated
- Aug 23, 2026