lordx64 · vision language · mixture of experts

Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled

lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled

Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled at Q2_K is exactly 13,246,129,568 bytes (12.34 GiB / 13.25 GB) — an effective 2.947 bits per weight, not the nominal 2. Its KV cache at 32K is 0.63 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
36.0B
total, not active
Architecture
qwen35moe
40 layers
Context
262,144
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
Q2_K12.34 GiB13,246,129,5682.947—huihui-ai
Q3_K15.99 GiB17,165,606,3043.820—huihui-ai
IQ4_XS17.64 GiB18,939,312,8964.214733lordx64
Q4_K20.22 GiB21,712,410,0164.831—huihui-ai
Q5_K_M23.03 GiB24,729,131,7765.503733lordx64
Q5_K23.61 GiB25,346,479,5205.640—huihui-ai
Q6_K27.20 GiB29,207,678,3686.499753huihui-ai
Q8_034.37 GiB36,903,140,0968.212733lordx64
Q8_035.21 GiB37,801,096,6088.412753huihui-ai
F1664.61 GiB69,376,638,04815.438733mudler
F1666.19 GiB71,065,941,72815.814753huihui-ai
F1666.19 GiB71,065,942,40015.814753mudler

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.08 GiB0.31 GiB4.00×10 / 0 / 30
8,1920.16 GiB0.63 GiB4.00×10 / 0 / 30
16,3840.31 GiB1.25 GiB4.00×10 / 0 / 30
32,7680.63 GiB2.50 GiB4.00×10 / 0 / 30
65,5361.25 GiB5.00 GiB4.00×10 / 0 / 30
131,0722.50 GiB10.00 GiB4.00×10 / 0 / 30

30 of 40 layers use linear attention, which keeps a fixed-size recurrent state instead of a per-token cache. Those layers do not grow with context at all — treating them as ordinary attention, as a flat formula does, overstates this model's cache by roughly 4.0× at long context.

Compare with

same modality, comparable size
vs Qwen3.6-35B-A3Bvs Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPvs Qwen3.5-9Bvs gemma-4-26B-A4B-itvs gemma-4-12B-itvs Qwen3.5-4BAll alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q2_K at roughly 18.83 GiB. The real file is 12.34 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
40
Attention heads
16
KV heads
2
Head dim
256
Hidden size
2048
Vocab
248,320
Sliding window
none
SWA period
—
MLA
no
Experts
256
Experts per token
8
use_sliding_window
—

Questions people ask

How much VRAM does Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled need?
Q2_K is exactly 13,246,129,568 bytes (12.34 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled's KV cache?
0.63 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled a mixture-of-experts model?
Yes — 256 experts, 8 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled should I use?
Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.