nvidia · text · mixture of experts

Nemotron-Cascade-2-30B-A3B

nvidia/Nemotron-Cascade-2-30B-A3B

Nemotron-Cascade-2-30B-A3B at Q4_K_M is exactly 24,725,740,992 bytes (23.03 GiB / 24.73 GB) — an effective 6.264 bits per weight, not the nominal 4. Its KV cache at 32K is 1.63 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
31.6B
total, not active
Architecture
nemotron_h_moe
52 layers
Context
262,144
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
I1-IQ1_S16.64 GiB17,864,524,6724.526—mradermacher
I1-IQ1_M16.64 GiB17,870,459,7764.527—mradermacher
I1-IQ2_XXS16.65 GiB17,880,351,6164.530—mradermacher
I1-IQ2_XS16.66 GiB17,890,329,4724.532—mradermacher
I1-IQ2_S16.67 GiB17,899,619,2004.535—mradermacher
I1-IQ2_M16.68 GiB17,907,532,6724.537—mradermacher
I1-Q2_K16.68 GiB17,908,994,9444.537—mradermacher
I1-IQ3_XXS16.69 GiB17,922,843,5204.541—mradermacher
I1-IQ3_XS16.70 GiB17,934,713,7284.544—mradermacher
I1-Q3_K_S16.70 GiB17,934,713,7284.544—mradermacher
I1-IQ3_S16.70 GiB17,934,713,7284.544—mradermacher
I1-IQ4_XS16.73 GiB17,967,141,7604.552—mradermacher
I1-Q4_016.82 GiB18,058,189,6964.575—mradermacher
I1-Q2_K_S16.82 GiB18,064,425,8564.577—mradermacher
IQ2_XXS16.84 GiB18,086,359,4884.582—bartowski
IQ2_XS16.85 GiB18,090,488,2564.583—bartowski
IQ2_S16.86 GiB18,098,745,7924.585—bartowski
IQ2_M16.86 GiB18,102,874,5604.586—bartowski
I1-IQ3_M16.86 GiB18,106,143,6164.587—mradermacher
Q2_K16.91 GiB18,161,537,4724.601—bartowski
IQ4_XS16.92 GiB18,166,698,4324.602—bartowski
IQ4_NL16.92 GiB18,170,827,2004.603—bartowski
Q3_K_S16.94 GiB18,189,922,7524.608—bartowski
Q4_017.02 GiB18,272,326,0804.629—bartowski
Q2_K_L17.08 GiB18,337,698,2404.646—bartowski
IQ3_XXS17.74 GiB19,046,814,1444.825—bartowski
IQ3_XS17.74 GiB19,053,007,2964.827—bartowski
Q3_K_M17.76 GiB19,067,974,0804.831—bartowski
Q3_K_L18.16 GiB19,498,226,1124.940—bartowski
IQ3_M18.16 GiB19,500,290,4964.940—bartowski
I1-Q3_K_M18.49 GiB19,849,236,3525.029—mradermacher
I1-Q4_118.56 GiB19,928,166,7845.049—mradermacher
Q4_118.74 GiB20,120,369,0885.097—bartowski
I1-Q3_K_L19.32 GiB20,749,436,8005.257—mradermacher
I1-Q4_K_S20.42 GiB21,920,544,6405.553—mradermacher
Q4_K_S20.91 GiB22,447,521,2165.687—bartowski
I1-Q5_K_S22.18 GiB23,810,305,4086.032—mradermacher
Q5_K_S22.36 GiB24,008,614,8486.082—bartowski
I1-Q4_K_M22.83 GiB24,515,131,2646.211—mradermacher
Q4_K_M23.03 GiB24,725,740,9926.264—bartowski

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.20 GiB0.20 GiB—52 / 0 / 0
8,1920.41 GiB0.41 GiB—52 / 0 / 0
16,3840.81 GiB0.81 GiB—52 / 0 / 0
32,7681.63 GiB1.63 GiB—52 / 0 / 0
65,5363.25 GiB3.25 GiB—52 / 0 / 0
131,0726.50 GiB6.50 GiB—52 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs gemma-4-E4B-itvs Qwen3-30B-A3B-Thinking-2507vs Qwen3-4BAll 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 Q4_K_M at roughly 16.54 GiB. The real file is 23.03 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
52
Attention heads
32
KV heads
2
Head dim
128
Hidden size
2688
Vocab
131,072
Sliding window
none
SWA period
—
MLA
no
Experts
128
Experts per token
6
use_sliding_window
—

Questions people ask

How much VRAM does Nemotron-Cascade-2-30B-A3B need?
Q4_K_M is exactly 24,725,740,992 bytes (23.03 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Nemotron-Cascade-2-30B-A3B's KV cache?
1.63 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 Nemotron-Cascade-2-30B-A3B a mixture-of-experts model?
Yes — 128 experts, 6 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 Nemotron-Cascade-2-30B-A3B 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.