ATH-MaaS · text

Marco-DeepResearch-8B

ATH-MaaS/Marco-DeepResearch-8B

Marco-DeepResearch-8B at Q4_K_M is exactly 5,027,784,128 bytes (4.68 GiB / 5.03 GB) — an effective 4.911 bits per weight, not the nominal 4. Its KV cache at 32K is 4.50 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
8.2B
Architecture
qwen3
36 layers
Context
131,072
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
I1-IQ1_S1.97 GiB2,115,770,0482.067—ATH-MaaS
I1-IQ1_S1.97 GiB2,115,770,4962.067—mradermacher
I1-IQ1_M2.10 GiB2,256,148,1602.204—ATH-MaaS
I1-IQ1_M2.10 GiB2,256,148,6082.204—mradermacher
I1-IQ2_XXS2.32 GiB2,490,111,6802.432—ATH-MaaS
I1-IQ2_XXS2.32 GiB2,490,112,1282.432—mradermacher
I1-IQ2_XS2.51 GiB2,696,156,8642.633—ATH-MaaS
I1-IQ2_XS2.51 GiB2,696,157,3122.633—mradermacher
I1-IQ2_S2.67 GiB2,864,744,1282.798—ATH-MaaS
I1-IQ2_S2.67 GiB2,864,744,5762.798—mradermacher
I1-IQ2_M2.84 GiB3,051,914,9442.981—ATH-MaaS
I1-IQ2_M2.84 GiB3,051,915,3922.981—mradermacher
I1-Q2_K_S2.87 GiB3,083,552,4483.012—ATH-MaaS
I1-Q2_K_S2.87 GiB3,083,552,8963.012—mradermacher
Q2_K3.06 GiB3,281,733,0563.205—ATH-MaaS
I1-Q2_K3.06 GiB3,281,733,3123.205—ATH-MaaS
I1-Q2_K3.06 GiB3,281,733,7603.205—mradermacher
I1-IQ3_XXS3.14 GiB3,369,633,4723.291—ATH-MaaS
I1-IQ3_XXS3.14 GiB3,369,633,9203.291—mradermacher
I1-IQ3_XS3.38 GiB3,626,874,5603.542—ATH-MaaS
I1-IQ3_XS3.38 GiB3,626,875,0083.542—mradermacher
Q3_K_S3.51 GiB3,769,611,7123.682—ATH-MaaS
I1-Q3_K_S3.51 GiB3,769,611,9683.682—ATH-MaaS
I1-Q3_K_S3.51 GiB3,769,612,4163.682—mradermacher
I1-IQ3_S3.53 GiB3,789,665,9843.701—ATH-MaaS
I1-IQ3_S3.53 GiB3,789,666,4323.701—mradermacher
I1-IQ3_M3.63 GiB3,896,620,7363.806—ATH-MaaS
I1-IQ3_M3.63 GiB3,896,621,1843.806—mradermacher
Q3_K_M3.84 GiB4,124,161,4724.028—ATH-MaaS
I1-Q3_K_M3.84 GiB4,124,161,7284.028—ATH-MaaS
I1-Q3_K_M3.84 GiB4,124,162,1764.028—mradermacher
Q3_K_L4.13 GiB4,431,394,2404.328—ATH-MaaS
I1-Q3_K_L4.13 GiB4,431,394,4964.328—ATH-MaaS
I1-Q3_K_L4.13 GiB4,431,394,9444.328—mradermacher
I1-IQ4_XS4.25 GiB4,561,839,8084.456—ATH-MaaS
I1-IQ4_XS4.25 GiB4,561,840,2564.456—mradermacher
IQ4_XS4.28 GiB4,593,296,8324.486—ATH-MaaS
I1-Q4_04.46 GiB4,787,332,8004.676—ATH-MaaS
I1-Q4_04.46 GiB4,787,333,2484.676—mradermacher
I1-IQ4_NL4.46 GiB4,793,624,2564.682—ATH-MaaS

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.56 GiB0.56 GiB—36 / 0 / 0
8,1921.13 GiB1.13 GiB—36 / 0 / 0
16,3842.25 GiB2.25 GiB—36 / 0 / 0
32,7684.50 GiB4.50 GiB—36 / 0 / 0
65,5369.00 GiB9.00 GiB—36 / 0 / 0
131,07218.00 GiB18.00 GiB—36 / 0 / 0

Compare with

same modality, comparable size
vs gemma-4-E4B-itvs Qwen3-4Bvs Qwen3-8Bvs HyperCLOVAX-SEED-Text-Instruct-1.5Bvs Llama-3.2-1B-Instructvs Qwen3-VL-8B-Instruct-abliterated-v1All 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 4.29 GiB. The real file is 4.68 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
36
Attention heads
32
KV heads
8
Head dim
128
Hidden size
4096
Vocab
151,936
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
false

Questions people ask

How much VRAM does Marco-DeepResearch-8B need?
Q4_K_M is exactly 5,027,784,128 bytes (4.68 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Marco-DeepResearch-8B's KV cache?
4.50 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.
Which quantization of Marco-DeepResearch-8B 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.