Salesforce · text · mixture of experts

xLAM-8x7b-r

Salesforce/xLAM-8x7b-r

xLAM-8x7b-r at Q4_K_M is exactly 28,448,467,936 bytes (26.49 GiB / 28.45 GB) — an effective 4.873 bits per weight, not the nominal 4. Its KV cache at 32K is 4.00 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
46.7B
total, not active
Architecture
llama
32 layers
Context
32,768
native (config.json)
License
cc-by-nc-4.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ1_S9.15 GiB9,821,670,3361.682—legraphista
IQ1_M10.10 GiB10,847,177,6641.858—legraphista
IQ2_XXS11.69 GiB12,556,356,5442.151—legraphista
IQ2_XS12.97 GiB13,923,699,6482.385—legraphista
IQ2_S13.16 GiB14,127,852,4802.420—legraphista
IQ2_S13.16 GiB14,127,852,5122.420—bartowski
IQ2_M14.43 GiB15,495,195,5842.654—legraphista
IQ2_M14.43 GiB15,495,195,6162.654—bartowski
Q2_K_S14.93 GiB16,031,968,1922.746—legraphista
Q2_K16.12 GiB17,311,230,9122.965—legraphista
Q2_K16.12 GiB17,311,230,9442.965—bartowski
Q2_K_L16.24 GiB17,439,230,9442.987—bartowski
IQ3_XXS16.99 GiB18,242,464,7043.125—legraphista
IQ3_XS18.02 GiB19,350,391,7443.315—legraphista
IQ3_XS18.02 GiB19,350,391,7763.315—bartowski
Q3_K_S19.03 GiB20,432,522,1763.500—legraphista
IQ3_S19.03 GiB20,432,522,1763.500—legraphista
Q3_K_S19.03 GiB20,432,522,2083.500—bartowski
IQ3_M19.96 GiB21,430,766,5283.671—legraphista
IQ3_M19.96 GiB21,430,766,5603.671—bartowski
Q3_K21.00 GiB22,546,451,3923.862—legraphista
Q3_K_M21.00 GiB22,546,451,4243.862—bartowski
Q3_K_L22.51 GiB24,169,647,0404.140—legraphista
Q3_K_L22.51 GiB24,169,647,0724.140—bartowski
IQ4_XS23.36 GiB25,080,540,0964.296—legraphista
IQ4_XS23.36 GiB25,080,540,1284.296—bartowski
IQ4_NL24.69 GiB26,510,699,4564.541—legraphista
Q4_024.74 GiB26,561,031,1364.550—bartowski
Q4_K_S24.91 GiB26,745,580,4804.581—legraphista
Q4_K_S24.91 GiB26,745,580,5124.581—bartowski
Q4_K26.49 GiB28,448,467,9044.873—legraphista
Q4_K_M26.49 GiB28,448,467,9364.873—bartowski
Q4_K_L26.59 GiB28,545,747,9364.890—bartowski
Q5_K_S30.02 GiB32,231,336,6725.521—legraphista
Q5_K_S30.02 GiB32,231,336,9285.521—bartowski
Q5_K30.95 GiB33,229,581,0245.692—legraphista
Q5_K_M30.95 GiB33,229,581,2805.692—bartowski
Q5_K_L31.02 GiB33,310,477,2805.706—bartowski
Q6_K35.74 GiB38,380,817,1206.574—legraphista
Q6_K35.74 GiB38,380,817,3766.574—bartowski

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.50 GiB0.50 GiB—32 / 0 / 0
8,1921.00 GiB1.00 GiB—32 / 0 / 0
16,3842.00 GiB2.00 GiB—32 / 0 / 0
32,7684.00 GiB4.00 GiB—32 / 0 / 0
65,5368.00 GiB8.00 GiB—32 / 0 / 0
131,07216.00 GiB16.00 GiB—32 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs gemma-4-12B-it-qat-q4_0-unquantizedvs gemma-4-E4B-itvs Qwen3-30B-A3B-Thinking-2507All 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 24.47 GiB. The real file is 26.49 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
32
Attention heads
32
KV heads
8
Head dim
128
Hidden size
4096
Vocab
32,000
Sliding window
none
SWA period
—
MLA
no
Experts
8
Experts per token
2
use_sliding_window
—

Questions people ask

How much VRAM does xLAM-8x7b-r need?
Q4_K_M is exactly 28,448,467,936 bytes (26.49 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is xLAM-8x7b-r's KV cache?
4.00 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 xLAM-8x7b-r a mixture-of-experts model?
Yes — 8 experts, 2 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 xLAM-8x7b-r 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.