How much VRAM does Hy3 need?

Q4_K_M is exactly 180,999,059,328 bytes (168.57 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Hy3's KV cache?

10.00 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

Is Hy3 a mixture-of-experts model?

Yes — 192 experts with 8 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

tencent · text · mixture of experts

Hy3

tencent/Hy3

Hy3 at Q4_K_M is exactly 180,999,059,328 bytes (168.57 GiB / 181.00 GB) — an effective 4.846 bits per weight, not the nominal 4. Its KV cache at 32K is 10.00 GiB.

From the file· summed from 4 file(s)From the file· KV per layer

Parameters

299B

total, not active

Architecture

hy_v3

80 layers

Context

262,144

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ1_S2 shards	59.65 GiB	64,051,148,352	1.715	—	bartowski
IQ1_M2 shards	66.22 GiB	71,100,724,800	1.904	—	bartowski
IQ2_XXS3 shards	76.47 GiB	82,113,918,656	2.199	—	bartowski
IQ1_M	83.30 GiB	89,446,312,384	2.395	—	AngelSlim
IQ2_XS3 shards	84.81 GiB	91,060,369,120	2.438	—	bartowski
IQ1_M2 shards	85.45 GiB	91,756,066,464	2.457	—	vcruz305
IQ2_S3 shards	86.43 GiB	92,805,134,016	2.485	—	bartowski
IQ2_M3 shards	93.14 GiB	100,008,834,816	2.678	—	vcruz305
IQ2_M3 shards	95.08 GiB	102,091,323,104	2.733	—	bartowski
Q2_K3 shards	99.36 GiB	106,691,229,376	2.857	—	bartowski
Q2_K_L3 shards	99.81 GiB	107,174,557,408	2.870	—	bartowski
Q2_K3 shards	101.28 GiB	108,748,601,120	2.912	—	vcruz305
IQ3_XXS3 shards	109.28 GiB	117,335,832,352	3.142	—	vcruz305
IQ3_XXS4 shards	117.43 GiB	126,089,361,248	3.376	—	bartowski
Q3_K_S4 shards	122.26 GiB	131,274,569,536	3.515	—	bartowski
Q3_K_M	125.35 GiB	134,597,913,024	3.604	1278	AngelSlim
IQ3_XS4 shards	128.00 GiB	137,437,050,656	3.680	—	bartowski
Q3_K_M4 shards	128.10 GiB	137,545,479,968	3.683	—	bartowski
Q3_K_M3 shards	132.94 GiB	142,740,092,672	3.822	—	vcruz305
Q3_K_L4 shards	133.25 GiB	143,077,766,976	3.831	—	bartowski
IQ3_M4 shards	133.80 GiB	143,668,639,584	3.847	—	bartowski
IQ4_XS4 shards	148.22 GiB	159,152,175,008	4.261	—	vcruz305
IQ4_XS5 shards	149.73 GiB	160,771,307,456	4.305	—	bartowski
IQ4_NL5 shards	158.14 GiB	169,803,872,192	4.546	—	bartowski
Q4_05 shards	158.62 GiB	170,321,082,336	4.560	—	bartowski
Q4_K_S5 shards	163.41 GiB	175,458,318,304	4.698	—	bartowski
Q4_K_M4 shards	168.57 GiB	180,999,059,328	4.846	—	vcruz305
Q4_K_M5 shards	169.65 GiB	182,158,227,360	4.877	—	bartowski
Q4_K_M	169.81 GiB	182,328,371,648	4.882	1278	AngelSlim
Q4_K_L5 shards	169.99 GiB	182,525,556,640	4.887	—	bartowski
Q4_15 shards	174.90 GiB	187,795,601,344	5.028	—	bartowski
Q5_K_S6 shards	191.88 GiB	206,034,401,344	5.517	—	bartowski
Q5_K_M5 shards	197.61 GiB	212,177,942,528	5.681	—	vcruz305
Q5_K_M6 shards	198.22 GiB	212,837,857,344	5.699	—	bartowski
Q6_K7 shards	239.58 GiB	257,246,623,936	6.888	—	bartowski
Q8_07 shards	295.84 GiB	317,652,367,296	8.505	—	vcruz305
Q8_09 shards	295.84 GiB	317,652,367,776	8.505	—	bartowski

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.25 GiB	1.25 GiB	—	80 / 0 / 0
8,192	2.50 GiB	2.50 GiB	—	80 / 0 / 0
16,384	5.00 GiB	5.00 GiB	—	80 / 0 / 0
32,768	10.00 GiB	10.00 GiB	—	80 / 0 / 0
65,536	20.00 GiB	20.00 GiB	—	80 / 0 / 0
131,072	40.00 GiB	40.00 GiB	—	80 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs DeepSeek-V4-Flash vs Qwen-AgentWorld-35B-A3B vs Llama-3.1-8B-Instruct vs Ornith-1.0-35B vs Qwen3-Coder-Next All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 156.52 GiB. The real file is 168.57 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

4096

Vocab

120,832

Sliding window

none

SWA period

—

MLA

Experts

192

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does Hy3 need?: Q4_K_M is exactly 180,999,059,328 bytes (168.57 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Hy3's KV cache?: 10.00 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 Hy3 a mixture-of-experts model?: Yes — 192 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 Hy3 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.