Brooooooklyn/Qwen3.5-35B-A3B-UD-Q8_K_XL-mlx

Qwen3.5-35B-A3B — UD-Q8_K_XL (mlx-node)

8-bit base mixed-precision quantization of Qwen/Qwen3.5-35B-A3B for Apple Silicon, using the Unsloth Dynamic quantization strategy via mlx-node.

	Original (BF16)	This Model
Size	~72 GB	35 GB
Format	SafeTensors (sharded)	SafeTensors (single file)
Precision	BF16 uniform	Mixed 8/8/8/8/8-bit + BF16

All Variants

Repo	GGUF Equivalent	Size	Decode (tok/s)	Speedup vs BF16
Brooooooklyn/Qwen3.5-35B-A3B-UD-Q2_K_XL-mlx	UD-Q2_K_XL	13 GB	86.1	2.1x
Brooooooklyn/Qwen3.5-35B-A3B-UD-Q3_K_XL-mlx	UD-Q3_K_XL	17 GB	79.7	2.0x
Brooooooklyn/Qwen3.5-35B-A3B-UD-Q4_K_XL-mlx	UD-Q4_K_XL	21 GB	73.4	1.8x
Brooooooklyn/Qwen3.5-35B-A3B-UD-Q5_K_XL-mlx	UD-Q5_K_XL	25 GB	70.0	1.7x
Brooooooklyn/Qwen3.5-35B-A3B-UD-Q6_K_XL-mlx	UD-Q6_K_XL	30 GB	64.9	1.6x
Brooooooklyn/Qwen3.5-35B-A3B-UD-Q8_K_XL-mlx	UD-Q8_K_XL	35 GB	63.5	1.6x

Benchmarked on Apple M3 Max 128GB, multi-turn chat (Turn 4 decode, steady-state).

Performance

Model	Size	Decode (tok/s)	Speedup
BF16 (unquantized)	65 GB	40.5	baseline
This model (UD-Q8_K_XL)	35 GB	63.5	1.6x faster

Decode is memory-bandwidth bound on Apple Silicon — fewer bytes per token directly translates to higher throughput. The MoE architecture activates only 8 of 256 experts per token (~3B active out of 35.9B total).

Per-Tensor Bit Assignments (N=8)

Weight	Bits	Rationale
embed_tokens	8-bit	KLD ~0.15 — very low sensitivity
lm_head	8-bit	KLD ~0.05 — safest tensor
self_attn.q/k/v_proj	8-bit + AWQ	KLD ~1.5–2.9, AWQ via layernorm
linear_attn.in_proj_qkv/z	8-bit + AWQ	KLD ~2.9, AWQ via layernorm
self_attn.o_proj	bf16	NOT AWQ-correctable
linear_attn.out_proj	bf16	KLD ~6.0 — worst tensor
down_proj	8-bit	"Slightly more sensitive"
gate_proj, up_proj	8-bit	"Generally ok" at low bits
Router gates	8-bit	MoE routing accuracy
GDN params (A_log, etc)	bf16	State-space dynamics

Quantization Strategy

Based on Unsloth Dynamic 2.0 per-tensor KLD analysis. Sensitive layers get higher bits with AWQ correction, while the bulk of FFN expert weights are aggressively quantized. imatrix AWQ pre-scaling amplifies important weight channels and fuses inverse scales into preceding layer norms (zero inference overhead).

AWQ-correctable projections (q/k/v, in_proj_qkv/z) are quantized at 8-bit via input_layernorm. Non-AWQ-correctable projections (o_proj, out_proj) are kept at bf16 — their inputs come from attention/GDN computation, not from a norm layer.

Architecture

Parameter	Value
Total parameters	35.9B (3B active per token)
Hidden size	2,048
Layers	40 (30 linear + 10 full attention)
Attention heads	16 (2 KV heads, GQA 8:1)
Head dimension	256
Experts	256 per MoE layer, top-8 routing
Vocab size	248,320
Max context	262,144 tokens

Usage

import { loadModel } from '@mlx-node/lm';

const model = await loadModel('./Qwen3.5-35B-A3B-UD-Q8_K_XL-mlx');

const result = await model.chat(
  [{ role: 'user', content: 'Explain the hybrid attention mechanism in Qwen3.5.' }],
  { maxNewTokens: 2048, temperature: 0.6, enableThinking: false },
);
console.log(result.text);

How It Was Made

mlx convert \
  -i Qwen3.5-35B-A3B \
  -o Qwen3.5-35B-A3B-UD-Q8_K_XL-mlx \
  -q --q-bits 8 --q-recipe unsloth \
  --imatrix-path imatrix_unsloth.gguf

Acknowledgments

• Unsloth — Quantization strategy based on their per-layer KLD benchmarks and Dynamic 2.0 methodology
• Qwen Team — For the Qwen3.5 model family
• Apple MLX — For the Metal-accelerated ML framework

License

Apache 2.0 (inherited from base model).