Open models, and what Glyd saves on each

The open models people run most, their size as released and with Glyd, and the smallest single GPU that holds them.

Updated Oct 5, 202622 measured · 32 estimated

Released in	Glyd saves, %	Which models
bf16	32 to 33	Most models you can run on one GPU: Qwen3.5, Gemma 4, Nemotron 3, Llama, Mistral Small. Every matrix, measured on twenty-one. The whole model in vLLM: Qwen3.5-9B 28.2% (RTX 4090) and 21.3% (RTX 5090), Qwen3.8-27B 22.8% (H100), parts of these models not compressed yet.
FP8	17 to 18	The measured floor on three FP8 models; lossless FP8 checkpoints (GLM-5.3, DeepSeek-V4, Mistral Small 4) are not shipped yet.
4-bit	≈ 0	gpt-oss, DeepSeek-V4.1 Flash, Kimi K3, MiMo-V2.6 Pro: already rounded to 4 bits, with little left to take.

As releasedWith Glyd

Run on one GPU

Model	Released	Format	As released	With Glyd	Saved, %	As released	With Glyd
			Size, GB			Smallest single GPU, GB
Qwen3.5-2Best.Qwen · 2.3B	Feb 2026	bf16	4.6	3.1	32.7	16	16
SmolLM3-3BHugging Face · 3.1B	Jul 2025	bf16	6.2	4.1	32.9	16	16
Llama-3.2-3B-InstructMeta · 3.2B	Sep 2024	bf16	6.4	4.3	32.8	16	16
NVIDIA-Nemotron-3-Nano-4B-BF16est.NVIDIA · 4B	Mar 2026	bf16	8.0	5.3	32.8	16	16
Qwen3-4B-Instruct-2507Qwen · 4B	Aug 2025	bf16	8.0	5.5	32.2	16	16
Qwen3.5-4Best.Qwen · 4.7B	Feb 2026	bf16	9.3	6.3	32.8	16	16
gemma-4-E2B-itest.Google · 5.1B	Mar 2026	bf16	10.3	6.9	32.8	16	16
gpt-oss-20best.OpenAI	Aug 2025	4-bit	13.8	12.6	8.6	16	16
Mistral-7B-Instruct-v0.3Mistral AI · 7.2B	May 2024	bf16	14.5	9.8	32.7	24	16
gemma-4-E4B-itest.Google · 8B	Mar 2026	bf16	16.0	10.8	32.6	24	16
Llama-3.1-8B-InstructMeta · 8B	Jul 2024	bf16	16.1	10.8	32.8	24	16
Qwen3-8BQwen · 8.2B	Apr 2025	bf16	16.4	11.2	31.9	24	16
MiMo-V2.6-Distill-Qwen-9Best.Xiaomi · 9.4B	Sep 2026	bf16	18.8	12.7	32.7	24	16
Qwen3.5-9BQwen · 9.7B	Feb 2026	bf16	19.3	13.9	28.2	24	16
gemma-4-12B-itest.Google · 12B	May 2026	bf16	23.9	16.1	32.7	32	24
gemma-3-12b-itGoogle · 12.2B	Mar 2025	bf16	24.4	16.4	32.9	32	24
phi-4Microsoft · 14.7B	Dec 2024	bf16	29.3	19.7	32.9	32	24
DeepSeek-R1-Distill-Qwen-14BDeepSeek · 14.8B	Jan 2025	bf16	29.5	20.1	31.9	32	24
Mistral-Small-3.2-24B-Instruct-2506Mistral AI · 24B	Jun 2025	bf16	48.0	32.2	33.0	48	48
gemma-4-26B-A4B-itGoogle · 25.8B	Mar 2026	bf16	51.6	34.7	32.8	80	48
gemma-3-27b-itGoogle · 27.4B	Mar 2025	bf16	54.9	36.9	32.8	80	48
Qwen3.8-27BQwen · 27.8B	Aug 2026	bf16	55.6	37.3	32.8	80	48
Qwen3.5-27Best.Qwen · 27.8B	Feb 2026	bf16	55.6	37.4	32.7	80	48
Qwen3.6-27Best.Qwen · 27.8B	Apr 2026	bf16	55.6	37.4	32.7	80	48
Muse-Glimmer-30BMeta · 29.8B	Aug 2026	bf16	59.6	40.0	32.8	80	48
Qwen3-30B-A3BQwen · 30.5B	Apr 2025	bf16	61.1	41.1	32.7	80	48
gemma-4-31B-itest.Google · 31.3B	Mar 2026	bf16	62.6	42.1	32.7	80	48
gpt-oss-120best.OpenAI	Aug 2025	4-bit	65.3	63.8	2.2	80	80
Qwen3-32BQwen · 32.8B	Apr 2025	bf16	65.5	44.5	32.1	80	48
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16est.NVIDIA · 32.9B	Aug 2026	bf16	65.8	44.3	32.7	80	48
Qwen3.5-35B-A3Best.Qwen · 36B	Feb 2026	bf16	71.9	48.4	32.7	80	48
Qwen3.6-35B-A3Best.Qwen · 36B	Apr 2026	bf16	71.9	48.4	32.7	80	48
Mistral-Small-4-119B-2603est.Mistral AI · 119.4B	Jan 2026	fp8	120.9	119.9	0.8	141	141
Llama-3.3-70B-InstructMeta · 70.6B	Nov 2024	bf16	141.1	94.7	32.9	141	96
Qwen2.5-72B-InstructQwen · 72.7B	Sep 2024	bf16	145.4	97.8	32.7	141	96
Qwen3-Coder-Nextest.Qwen · 79.7B	Jan 2026	bf16	159.4	107.3	32.7	—	141
Qwen3-Next-80B-A3B-InstructQwen · 81.3B	Sep 2025	bf16	162.7	110.3	32.2	—	141
Llama-4-Scout-17B-16E-InstructMeta · 108.6B	Apr 2025	bf16	217.3	145.8	32.9	—	141
GLM-4.5-AirZ.ai · 110.5B	Jul 2025	bf16	220.9	148.0	33.0	—	—
Qwen3.5-122B-A10Best.Qwen · 125.1B	Feb 2026	bf16	250.2	168.4	32.7	—	—

Frontier models

Model	Released	Format	As released	With Glyd	Saved, %	As released	With Glyd
			Size, GB			8×H100 nodes
Mistral-Medium-3.5-128Best.Mistral AI · 127.7B	Mar 2026	fp8	133.6	129.8	2.9	1	1
GLM-5.3-Flashest.Z.ai · 321.3B	Aug 2026	fp8	328.3	323.8	1.4	1	1
DeepSeek-V4.1-Flashest.DeepSeek	Sep 2026	4-bit	510.3	509.0	0.3	1	1
MiMo-V2.6-Pro-RLest.Xiaomi	Sep 2026	4-bit	566.1	559.1	1.2	1	1
GLM-5.3-Flash-BF16est.Z.ai · 321.3B	Aug 2026	bf16	642.7	432.5	32.7	1	1
GLM-5.3est.Z.ai · 753.3B	Aug 2026	fp8	755.6	754.3	0.2	2	2
K2-Horizon-375B-A23Best.IFM · 379.2B	Sep 2026	bf16	758.3	510.4	32.7	2	1
Qwen3.5-397B-A17Best.Qwen · 403.4B	Feb 2026	bf16	806.8	543.0	32.7	2	1
MiniMax-M3est.MiniMax · 427B	Jun 2026	bf16	854.2	574.9	32.7	2	1
DeepSeek-V4-Proest.DeepSeek	Apr 2026	4-bit	864.7	862.9	0.2	2	2
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16est.NVIDIA · 560.5B	Jun 2026	bf16	1,121.1	754.5	32.7	2	2
GLM-5.3-BF16est.Z.ai · 753.3B	Aug 2026	bf16	1,506.7	1,014.0	32.7	3	2
Kimi-K3est.Moonshot AI	Jun 2026	4-bit	1,560.9	1,523.5	2.4	3	3
Inklingest.Thinking Machines · 952.4B	Jul 2026	bf16	1,904.8	1,281.9	32.7	3	2

Sizes are the checkpoints on Hugging Face; est.: worked out from the parameter count at the measured 32.7% for bf16 (the smallest layout). A measured row is every matrix packed on an A10 or an H100 (the 2025 set, Qwen3.8 27B, Gemma 4 26B-A4B and Muse Glimmer 30B), or, for Qwen3.5 9B, the whole model's weights in vLLM on an RTX 4090; each model's page has its rows. FP8 and 4-bit weights are counted as they are. A single GPU holds the weights, an 8K-token KV cache and 1.5 GB for the runtime; —: none does.