cerebras · text · mixture of experts

GLM-4.6-REAP-268B-A32B

cerebras/GLM-4.6-REAP-268B-A32B

GLM-4.6-REAP-268B-A32B at Q4_K_M is exactly 162,527,234,336 bytes (151.37 GiB / 162.53 GB) — an effective 4.837 bits per weight, not the nominal 4. Its KV cache at 32K is 11.50 GiB.

From the file· summed from 4 file(s)From the file· KV per layer
Parameters
269B
total, not active
Architecture
glm4moe
92 layers
Context
202,752
native (config.json)
License
mit

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
UD-TQ1_060.36 GiB64,814,184,3201.929—unsloth
UD-IQ1_S2 shards69.09 GiB74,185,013,3442.208—unsloth
UD-IQ1_M2 shards76.78 GiB82,440,910,9122.454—unsloth
UD-IQ2_XXS2 shards83.03 GiB89,148,889,1842.653—unsloth
UD-IQ2_M2 shards87.26 GiB93,690,533,9842.788—unsloth
Q2_K2 shards91.94 GiB98,719,828,0322.938—unsloth
Q2_K_L2 shards92.11 GiB98,901,690,4642.944—unsloth
UD-IQ3_XXS3 shards102.71 GiB110,280,849,6003.282—unsloth
Q3_K_S3 shards108.47 GiB116,470,417,6323.466—unsloth
Q3_K_M3 shards119.90 GiB128,744,163,5523.832—unsloth
IQ4_XS3 shards133.87 GiB143,746,009,3124.278—unsloth
IQ4_NL4 shards141.57 GiB152,014,686,5284.524—unsloth
Q4_04 shards141.97 GiB152,438,049,0884.537—unsloth
Q4_K_S4 shards142.42 GiB152,924,326,2084.551—unsloth
Q4_K_M4 shards151.37 GiB162,527,234,3364.837—unsloth
Q4_14 shards156.99 GiB168,570,390,8165.017—unsloth
Q5_K_S4 shards172.64 GiB185,372,305,7285.517—unsloth
Q5_K_M4 shards177.71 GiB190,818,920,8005.679—unsloth
Q6_K5 shards205.68 GiB220,850,813,3446.573—unsloth
Q8_06 shards266.14 GiB285,764,754,0168.505—unsloth
BF1611 shards500.77 GiB537,700,703,20016.004—unsloth

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0961.44 GiB1.44 GiB—92 / 0 / 0
8,1922.88 GiB2.88 GiB—92 / 0 / 0
16,3845.75 GiB5.75 GiB—92 / 0 / 0
32,76811.50 GiB11.50 GiB—92 / 0 / 0
65,53623.00 GiB23.00 GiB—92 / 0 / 0
131,07246.00 GiB46.00 GiB—92 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs DeepSeek-V4-Flashvs Qwen3-30B-A3B-Thinking-2507vs Laguna-XS-2.1All 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 Q4_K_M at roughly 140.81 GiB. The real file is 151.37 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
92
Attention heads
96
KV heads
8
Head dim
128
Hidden size
5120
Vocab
151,552
Sliding window
none
SWA period
—
MLA
no
Experts
120
Experts per token
8
use_sliding_window
—

Questions people ask

How much VRAM does GLM-4.6-REAP-268B-A32B need?
Q4_K_M is exactly 162,527,234,336 bytes (151.37 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is GLM-4.6-REAP-268B-A32B's KV cache?
11.50 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 GLM-4.6-REAP-268B-A32B a mixture-of-experts model?
Yes — 120 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 GLM-4.6-REAP-268B-A32B 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.