Where do these recipes come from?

From the vLLM recipes project (vllm-project/recipes) and NVIDIA Dynamo's recipes directory, both Apache 2.0. The control plane ships a snapshot of each at a pinned commit; this directory is built from the same snapshot, captured 2026-09-01.

Are the memory figures measured on Wingsmith?

No. Each memory floor is the minimum the vLLM recipe publishes for that precision, across the whole replica. The control plane shows it next to its own sizing and says so when the two disagree.

Can Wingsmith serve a model that is not listed here?

Yes. Any HuggingFace slug can be registered. A recipe adds published flags and a memory floor to the draft when one exists; without one, the plane sizes the weights from their own config.

Model directory

Serving recipes for
178 open models.

Every model in the vLLM recipe corpus that ships inside the control plane: tested precisions, published memory floors, parallelism layouts, and NVIDIA Dynamo recipes where NVIDIA publishes them. Register one of these slugs and its recipe comes with the draft.

Models: 178One page each, from the vLLM recipe corpus.
Providers: 41As the recipes name them.
Dynamo recipes: 18 modelsWith at least one NVIDIA Dynamo deployment recipe.
Snapshot: 2026-09-01vLLM recipes at 4bcc189.

Note

Memory floors and hardware are what the upstream recipes publish from their own runs, not measurements on your clusters. Floors are for the whole replica, not per GPU.

Qwen 32

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Qwen-ImageQwen/Qwen-Image	20B	Dense	Not stated	BF16, FP8, INT8	24 GB	None
Qwen2.5-32BQwen/Qwen2.5-32B	32B	Dense	128K tokens	BF16	77 GB	None
Qwen2.5-VL-72B-InstructQwen/Qwen2.5-VL-72B-Instruct	72B	Dense	125K tokens	BF16, INT4	43 GB	None
Qwen2.5-VL-7B-InstructQwen/Qwen2.5-VL-7B-Instruct	7B	Dense	125K tokens	BF16, INT4	5 GB	None
Qwen3-1.7BQwen/Qwen3-1.7B	1.7B	Dense	40K tokens	BF16	5 GB	None
Qwen3-14BQwen/Qwen3-14B	14B	Dense	40K tokens	BF16, FP8, INT4	10 GB	None
Qwen3-235B-A22B-InstructQwen/Qwen3-235B-A22B-Instruct-2507	235B total, 22B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	141 GB	TensorRT-LLM
Qwen3-30B-A3BQwen/Qwen3-30B-A3B	30.5B total, 3.3B active	Mixture of experts	40K tokens	BF16	74 GB	None
Qwen3-32BQwen/Qwen3-32B	32B	Dense	40K tokens	BF16, FP8, INT4	20 GB	vLLM, TensorRT-LLM
Qwen3-4BQwen/Qwen3-4B	4B	Dense	40K tokens	BF16, FP8	5 GB	None
Qwen3-8BQwen/Qwen3-8B	8.2B	Dense	40K tokens	BF16	20 GB	None
Qwen3-ASR-1.7BQwen/Qwen3-ASR-1.7B	2.3B	Dense	64K tokens	BF16	4 GB	None
Qwen3-Coder-480B-A35B-InstructQwen/Qwen3-Coder-480B-A35B-Instruct	480B total, 35B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	288 GB	None
Qwen3-Next-80B-A3B-InstructQwen/Qwen3-Next-80B-A3B-Instruct	80B total, 3B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	48 GB	None
Qwen3-TTS-12Hz-1.7B-CustomVoiceQwen/Qwen3-TTS-12Hz-1.7B-CustomVoice	1.7B	Dense	Not stated	BF16	8 GB	None
Qwen3-VL-235B-A22B-InstructQwen/Qwen3-VL-235B-A22B-Instruct	235B total, 22B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	141 GB	None
Qwen3-VL-30B-A3B-InstructQwen/Qwen3-VL-30B-A3B-Instruct	30B total, 3B active	Mixture of experts	256K tokens	BF16	72 GB	None
Qwen3.5-0.8BQwen/Qwen3.5-0.8B	0.8B	Dense	256K tokens	BF16	2 GB	None
Qwen3.5-122B-A10BQwen/Qwen3.5-122B-A10B	122B total, 10B active	Mixture of experts	256K tokens	BF16, FP8, INT4	74 GB	vLLM
Qwen3.5-27BQwen/Qwen3.5-27B	27B	Dense	256K tokens	BF16, FP8, INT4	17 GB	None
Qwen3.5-2BQwen/Qwen3.5-2B	2B	Dense	256K tokens	BF16	5 GB	None
Qwen3.5-35B-A3BQwen/Qwen3.5-35B-A3B	35B total, 3B active	Mixture of experts	256K tokens	BF16, FP8, INT4	21 GB	None
Qwen3.5-397BQwen/Qwen3.5-397B-A17B	397B total, 17B active	Mixture of experts	256K tokens	BF16, NVFP4, INT4	238 GB	None
Qwen3.5-4BQwen/Qwen3.5-4B	4B	Dense	256K tokens	BF16	10 GB	None
Qwen3.5-9BQwen/Qwen3.5-9B	9B	Dense	256K tokens	BF16	22 GB	None
Qwen3.6-27BQwen/Qwen3.6-27B	27B	Dense	256K tokens	BF16, FP8, NVFP4	27 GB	None
Qwen3.6-35B-A3BQwen/Qwen3.6-35B-A3B	35B total, 3B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	21 GB	vLLM
Qwen3.8-2.4T-A95BQwen/Qwen3.8-2.4T-A95B	2.4T total, 95B active	Mixture of experts	256K tokens	BF16, FP8, MXFP4, NVFP4	1373 GB	vLLM, SGLang
Qwen3.8-27BQwen/Qwen3.8-27B	27B	Dense	256K tokens	BF16, FP8, NVFP4, INT4, INT8	24 GB	None
Qwen3.8-Flash-NextQwen/Qwen3.8-Flash-Next	176B total, 6B active	Mixture of experts	256K tokens	NVFP4, FP8, BF16	130 GB	None
Qwen3Guard-Gen-8BQwen/Qwen3Guard-Gen-8B	8B	Dense	32K tokens	BF16	4 GB	None
QwQ-32BQwen/QwQ-32B	32B	Dense	40K tokens	BF16, INT8, INT4	20 GB	None

GLM (Z-AI) 16

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
GLM-4.5zai-org/GLM-4.5	358B total, 32B active	Mixture of experts	128K tokens	BF16, FP8	430 GB	None
GLM-4.5Vzai-org/GLM-4.5V	107B total, 12B active	Mixture of experts	64K tokens	BF16, FP8	128 GB	None
GLM-4.6zai-org/GLM-4.6	357B total, 32B active	Mixture of experts	198K tokens	BF16, FP8	428 GB	None
GLM-4.6Vzai-org/GLM-4.6V	107B total, 12B active	Mixture of experts	128K tokens	BF16, FP8	128 GB	None
GLM-4.7zai-org/GLM-4.7	358B total, 32B active	Mixture of experts	198K tokens	BF16, FP8, NVFP4	215 GB	None
GLM-5zai-org/GLM-5	744B total, 40B active	Mixture of experts	198K tokens	BF16, FP8, NVFP4	446 GB	SGLang
GLM-5.1zai-org/GLM-5.1	744B total, 40B active	Mixture of experts	198K tokens	BF16, FP8, NVFP4, MXFP4	446 GB	None
GLM-5.2zai-org/GLM-5.2	743B total, 39B active	Mixture of experts	1M tokens	FP8, MXFP4, NVFP4, BF16	446 GB	None
GLM-5.3zai-org/GLM-5.3	743B total, 39B active	Mixture of experts	1M tokens	FP8, NVFP4, BF16	558 GB	None
GLM-5.3-Flashzai-org/GLM-5.3-Flash	321B total, 18B active	Mixture of experts	1M tokens	FP8, NVFP4, BF16	229 GB	vLLM
GLM-ASR-Nano-2512zai-org/GLM-ASR-Nano-2512	2.3B	Dense	8K tokens	BF16	11 GB	None
GLM-GAzai-org/GLM-GA	10B	Dense	128K tokens	BF16	29 GB	None
GLM-Imagezai-org/GLM-Image	16B	Dense	4K tokens	BF16	38 GB	None
GLM-OCRzai-org/GLM-OCR	0.9B	Dense	128K tokens	BF16	2 GB	None
GLM-TTSzai-org/GLM-TTS	3B	Dense	Not stated	BF16	24 GB	None
Glyphzai-org/Glyph	10B	Dense	128K tokens	BF16	24 GB	None

NVIDIA 12

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Cosmos3-Nanonvidia/Cosmos3-Nano	16B	Dense	256K tokens	BF16	39 GB	None
Cosmos3-Supernvidia/Cosmos3-Super	64B	Dense	256K tokens	BF16	154 GB	None
Cosmos3-Super-Image2Videonvidia/Cosmos3-Super-Image2Video	64B	Dense	256K tokens	BF16	154 GB	None
Cosmos3-Super-Text2Imagenvidia/Cosmos3-Super-Text2Image	64B	Dense	256K tokens	BF16	154 GB	None
Nemotron-3-Nano-Omni-30B-A3B-Reasoningnvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16	31B total, 3B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	28 GB	vLLM
NVIDIA Nemotron 3.5 Lightningnvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16	30B total, 3B active	Mixture of experts	1M tokens	NVFP4, BF16	18 GB	vLLM, TensorRT-LLM
NVIDIA Nemotron-3-Nano-30B-A3Bnvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16	30B total, 3B active	Mixture of experts	256K tokens	BF16, FP8	35 GB	None
NVIDIA Nemotron-3-Nano-4Bnvidia/NVIDIA-Nemotron-3-Nano-4B-BF16	4B	Dense	256K tokens	BF16, FP8	5 GB	None
NVIDIA Nemotron-3-Super-120B-A12Bnvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	120B total, 12B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	75 GB	vLLM, SGLang, TensorRT-LLM
NVIDIA Nemotron-3-Ultra-550B-A55Bnvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16	550B total, 55B active	Mixture of experts	256K tokens	BF16, NVFP4	330 GB	vLLM
NVIDIA Nemotron-Nano-12B-v2-VLnvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16	12B	Dense	128K tokens	BF16, FP8, NVFP4	8 GB	None
NVIDIA Nemotron-Nano-9B-v2nvidia/NVIDIA-Nemotron-Nano-9B-v2	9B	Dense	128K tokens	BF16, FP8, NVFP4	6 GB	None

LiquidAI 10

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
LFM2.5 1.2B BaseLiquidAI/LFM2.5-1.2B-Base	1.2B	Dense	32K tokens	BF16	3 GB	None
LFM2.5 1.2B InstructLiquidAI/LFM2.5-1.2B-Instruct	1.2B	Dense	32K tokens	BF16	3 GB	None
LFM2.5 1.2B JPLiquidAI/LFM2.5-1.2B-JP	1.2B	Dense	32K tokens	BF16	3 GB	None
LFM2.5 1.2B JP (202606)LiquidAI/LFM2.5-1.2B-JP-202606	1.2B	Dense	32K tokens	BF16	3 GB	None
LFM2.5 1.2B ThinkingLiquidAI/LFM2.5-1.2B-Thinking	1.2B	Dense	32K tokens	BF16	3 GB	None
LFM2.5 230MLiquidAI/LFM2.5-230M	230M	Dense	32K tokens	BF16	1 GB	None
LFM2.5 350MLiquidAI/LFM2.5-350M	350M	Dense	32K tokens	BF16	1 GB	None
LFM2.5 8B A1BLiquidAI/LFM2.5-8B-A1B	8B total, 1B active	Mixture of experts	125K tokens	BF16	21 GB	None
LFM2.5 VL 1.6BLiquidAI/LFM2.5-VL-1.6B	1.6B	Dense	32K tokens	BF16	4 GB	None
LFM2.5 VL 450MLiquidAI/LFM2.5-VL-450M	450M	Dense	32K tokens	BF16	2 GB	None

DeepSeek 9

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
DeepSeek-OCRdeepseek-ai/DeepSeek-OCR	3B	Dense	8K tokens	BF16	7 GB	None
DeepSeek-OCR-2deepseek-ai/DeepSeek-OCR-2	3B	Dense	8K tokens	BF16	7 GB	None
DeepSeek-R1deepseek-ai/DeepSeek-R1	671B total, 37B active	Mixture of experts	160K tokens	FP8, NVFP4	403 GB	SGLang, TensorRT-LLM
DeepSeek-V3deepseek-ai/DeepSeek-V3	671B total, 37B active	Mixture of experts	160K tokens	FP8, FP4	403 GB	None
DeepSeek-V3.1deepseek-ai/DeepSeek-V3.1	671B total, 37B active	Mixture of experts	160K tokens	FP8, NVFP4	403 GB	None
DeepSeek-V3.2deepseek-ai/DeepSeek-V3.2	671B total, 37B active	Mixture of experts	160K tokens	FP8, NVFP4	403 GB	None
DeepSeek-V3.2-Expdeepseek-ai/DeepSeek-V3.2-Exp	671B total, 37B active	Mixture of experts	160K tokens	FP8	805 GB	None
DeepSeek-V4-Flashdeepseek-ai/DeepSeek-V4-Flash	284B total, 13B active	Mixture of experts	1M tokens	FP8, NVFP4	170 GB	vLLM, SGLang
DeepSeek-V4-Prodeepseek-ai/DeepSeek-V4-Pro	1.6T total, 49B active	Mixture of experts	1M tokens	FP8, NVFP4	960 GB	vLLM, SGLang

Google 7

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
DiffusionGemma 26B-A4B ITGoogle/diffusiongemma-26B-A4B-it	26B total, 4B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	24 GB	None
Gemma 4 12B ITGoogle/gemma-4-12B-it	12B	Dense	128K tokens	BF16, NVFP4, INT4, FP8	8 GB	None
Gemma 4 26B-A4B ITGoogle/gemma-4-26B-A4B-it	26B total, 4B active	Mixture of experts	128K tokens	BF16, FP8, NVFP4	16 GB	None
Gemma 4 31B ITGoogle/gemma-4-31B-it	31B	Dense	256K tokens	BF16, FP8, NVFP4, INT4	19 GB	None
Gemma 4 E2B ITGoogle/gemma-4-E2B-it	5B	Dense	128K tokens	BF16, FP8, INT4, INT2/4/8	4 GB	None
Gemma 4 E4B ITGoogle/gemma-4-E4B-it	8B	Dense	128K tokens	BF16, FP8, INT4, INT2/4/8	5 GB	None
TranslateGemma 27B ITGoogle/translategemma-27b-it	27B	Dense	128K tokens	BF16	10 GB	None

inclusionAI 7

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Ling-2.6-1TinclusionAI/Ling-2.6-1T	1T total, 50B active	Mixture of experts	256K tokens	FP8	1200 GB	None
Ling-2.6-flashinclusionAI/Ling-2.6-flash	104B total, 7.4B active	Mixture of experts	128K tokens	BF16	250 GB	None
Ling-3.0-flashinclusionAI/Ling-3.0-flash	124B total, 5.5B active	Mixture of experts	256K tokens	BF16, FP8, FP4, INT4	85 GB	None
Ling-3.0-tinyinclusionAI/Ling-3.0-tiny	7.9B total, 1.3B active	Mixture of experts	128K tokens	BF16, FP8, INT4	8 GB	None
Ming-omni-tts-0.5BinclusionAI/Ming-omni-tts-0.5B	0.5B	Dense	Not stated	BF16	16 GB	None
Ring-1T-FP8inclusionAI/Ring-1T-FP8	1T total, 50B active	Mixture of experts	64K tokens	FP8	1200 GB	None
Ring-2.6-1TinclusionAI/Ring-2.6-1T	1T total, 50B active	Mixture of experts	128K tokens	FP8	1200 GB	None

Meta 7

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Llama-3.1-8Bmeta-llama/Llama-3.1-8B	8B	Dense	128K tokens	BF16, INT8	10 GB	None
Llama-3.1-8B-Instructmeta-llama/Llama-3.1-8B-Instruct	8B	Dense	128K tokens	BF16, INT8, INT4, NVFP4, FP8	5 GB	None
Llama-3.2-1Bmeta-llama/Llama-3.2-1B	1.23B	Dense	128K tokens	BF16	3 GB	None
Llama-3.2-1B-Instructmeta-llama/Llama-3.2-1B-Instruct	1.23B	Dense	128K tokens	BF16, INT8, INT4	2 GB	None
Llama-3.2-3B-Instructmeta-llama/Llama-3.2-3B-Instruct	3.21B	Dense	128K tokens	BF16, INT8, INT4	3 GB	None
Llama-3.3-70Bmeta-llama/Llama-3.3-70B-Instruct	70B	Dense	128K tokens	BF16, FP8, NVFP4	42 GB	None
Llama-4-Scoutmeta-llama/Llama-4-Scout-17B-16E-Instruct	109B total, 17B active	Mixture of experts	10M tokens	BF16, FP8, NVFP4	65 GB	None

Mistral AI 7

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Ministral-3-Instructmistralai/Ministral-3-14B-Instruct-2512	14B	Dense	256K tokens	FP8	6 GB	None
Ministral-3-Reasoningmistralai/Ministral-3-8B-Reasoning-2512	8B	Dense	256K tokens	BF16	8 GB	None
Mistral-Large-3-675B-Instructmistralai/Mistral-Large-3-675B-Instruct-2512	675B total, 22B active	Mixture of experts	288K tokens	FP8, NVFP4	405 GB	None
Mistral-Medium-3.5mistralai/Mistral-Medium-3.5-128B	128B	Dense	256K tokens	FP8	154 GB	None
Mistral-Small-4-119Bmistralai/Mistral-Small-4-119B-2603	119B total, 6.5B active	Mixture of experts	256K tokens	FP8, NVFP4	72 GB	None
Voxtral-4B-TTS-2603mistralai/Voxtral-4B-TTS-2603	4B	Dense	Not stated	BF16	24 GB	None
Voxtral-Mini-4B-Realtime-2602mistralai/Voxtral-Mini-4B-Realtime-2602	4.4B	Dense	128K tokens	BF16	16 GB	None

Moonshot AI 7

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Kimi-K2-Instructmoonshotai/Kimi-K2-Instruct	1T total, 32B active	Mixture of experts	128K tokens	FP8	1200 GB	None
Kimi-K2-Thinkingmoonshotai/Kimi-K2-Thinking	1T total, 32B active	Mixture of experts	256K tokens	INT4, NVFP4	600 GB	None
Kimi-K2.5moonshotai/Kimi-K2.5	1T total, 32B active	Mixture of experts	256K tokens	INT4, NVFP4, MXFP4	600 GB	TensorRT-LLM
Kimi-K2.6moonshotai/Kimi-K2.6	1T total, 32B active	Mixture of experts	256K tokens	INT4, NVFP4	600 GB	vLLM
Kimi-K2.7-Codemoonshotai/Kimi-K2.7-Code	1T total, 32B active	Mixture of experts	256K tokens	INT4	714 GB	None
Kimi-K3moonshotai/Kimi-K3	2.8T total, 16 experts/token + shared (of 896 routed) active	Mixture of experts	1M tokens	MXFP4, NVFP4, INT4	1650 GB	vLLM
Kimi-Linear-48B-A3B-Instructmoonshotai/Kimi-Linear-48B-A3B-Instruct	48B total, 3B active	Mixture of experts	1M tokens	BF16	115 GB	None

MiniMax 6

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
MiniMax-H3MiniMaxAI/MiniMax-H3	64B	Dense	Not stated	BF16	202 GB	None
MiniMax-M2MiniMaxAI/MiniMax-M2	230B total, 10B active	Mixture of experts	192K tokens	FP8, NVFP4	138 GB	None
MiniMax-M2.1MiniMaxAI/MiniMax-M2.1	230B total, 10B active	Mixture of experts	192K tokens	FP8	276 GB	None
MiniMax-M2.5MiniMaxAI/MiniMax-M2.5	230B total, 10B active	Mixture of experts	192K tokens	FP8, NVFP4	138 GB	None
MiniMax-M2.7MiniMaxAI/MiniMax-M2.7	230B total, 10B active	Mixture of experts	192K tokens	FP8, NVFP4	138 GB	None
MiniMax-M3MiniMaxAI/MiniMax-M3	427B total, 26B active	Mixture of experts	1M tokens	BF16, MXFP8, NVFP4, MXFP4	257 GB	None

OpenMOSS 6

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
MOSS-SoundEffectOpenMOSS-Team/MOSS-SoundEffect	8B	Dense	Not stated	BF16	32 GB	None
MOSS-Transcribe-DiarizeOpenMOSS-Team/MOSS-Transcribe-Diarize	0.9B	Dense	Not stated	BF16	4 GB	None
MOSS-TTSOpenMOSS-Team/MOSS-TTS	8B	Dense	Not stated	BF16	32 GB	None
MOSS-TTS-RealtimeOpenMOSS-Team/MOSS-TTS-Realtime	1.7B	Dense	Not stated	BF16	16 GB	None
MOSS-TTSD-v1.0OpenMOSS-Team/MOSS-TTSD-v1.0	8B	Dense	Not stated	BF16	32 GB	None
MOSS-VoiceGeneratorOpenMOSS-Team/MOSS-VoiceGenerator	1.7B	Dense	Not stated	BF16	16 GB	None

Hunyuan (Tencent) 4

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Hunyuan-A13B-Instructtencent/Hunyuan-A13B-Instruct	80B total, 13B active	Mixture of experts	32K tokens	BF16, FP8	96 GB	None
HunyuanOCRtencent/HunyuanOCR	1B	Dense	32K tokens	BF16	2 GB	None
Hy3-previewtencent/Hy3-preview	295B total, 21B active	Mixture of experts	256K tokens	BF16	708 GB	None
Hy4-previewtencent/Hy4-preview	770B total, 49B active	Mixture of experts	1M tokens	BF16, MXFP8	924 GB	None

Poolside 4

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Laguna M.1poolside/Laguna-M.1	225B total, 23B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	135 GB	None
Laguna S-2.1poolside/Laguna-S-2.1	118B total, 8B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4, INT4	86 GB	None
Laguna XS-2.1poolside/Laguna-XS-2.1	33B total, 3B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4, INT4	26 GB	None
Laguna XS.2poolside/Laguna-XS.2	33B total, 3B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4, INT4	26 GB	None

Ernie (Baidu) 3

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
ERNIE-4.5baidu/ERNIE-4.5-21B-A3B-PT	21B total, 3B active	Mixture of experts	128K tokens	BF16	106 GB	None
ERNIE-4.5-VLbaidu/ERNIE-4.5-VL-28B-A3B-PT	28B total, 3B active	Mixture of experts	128K tokens	BF16, FP8	67 GB	None
Unlimited-OCRbaidu/Unlimited-OCR	3B	Dense	32K tokens	BF16	8 GB	None

Microsoft 3

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Phi-4microsoft/Phi-4-mini-instruct	4B	Dense	128K tokens	BF16	10 GB	None
Phi-4-multimodal-instructmicrosoft/Phi-4-multimodal-instruct	5.6B	Dense	128K tokens	BF16	14 GB	None
Phi-4-reasoningmicrosoft/Phi-4-reasoning	14B	Dense	32K tokens	BF16	34 GB	None

MiMo (Xiaomi) 3

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
MiMo-V2-FlashXiaomiMiMo/MiMo-V2-Flash	309B total, 15B active	Mixture of experts	256K tokens	FP8	371 GB	None
MiMo-V2.5XiaomiMiMo/MiMo-V2.5	311B total, 15B active	Mixture of experts	1M tokens	FP8	374 GB	None
MiMo-V2.5-ProXiaomiMiMo/MiMo-V2.5-Pro	1T total, 42B active	Mixture of experts	1M tokens	FP8	1224 GB	None

MiniCPM (OpenBMB) 3

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
MiniCPM-V 4.6openbmb/MiniCPM-V-4.6	1.3B	Dense	256K tokens	BF16	4 GB	None
MiniCPM5-1Bopenbmb/MiniCPM5-1B	1.1B	Dense	128K tokens	BF16	3 GB	None
VoxCPM2openbmb/VoxCPM2	2B	Dense	Not stated	BF16	24 GB	None

OpenAI 3

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
GPT-OSSopenai/gpt-oss-120b	120B total, 5.1B active	Mixture of experts	128K tokens	MXFP4	80 GB	vLLM, TensorRT-LLM
GPT-OSS 20Bopenai/gpt-oss-20b	21B total, 3.6B active	Mixture of experts	128K tokens	MXFP4	16 GB	None
Whisper Large v3openai/whisper-large-v3	1.55B	Dense	448 tokens	BF16	4 GB	None

PaddlePaddle 3

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
PaddleOCR-VLPaddlePaddle/PaddleOCR-VL	0.9B	Dense	128K tokens	BF16	2 GB	None
PaddleOCR-VL-1.5PaddlePaddle/PaddleOCR-VL-1.5	0.9B	Dense	128K tokens	BF16	2 GB	None
PaddleOCR-VL-1.6PaddlePaddle/PaddleOCR-VL-1.6	0.9B	Dense	128K tokens	BF16	2 GB	None

InternLM 2

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Intern-S1internlm/Intern-S1	241B total, 28B active	Mixture of experts	64K tokens	BF16, FP8	289 GB	None
Intern-S2-Previewinternlm/Intern-S2-Preview	36B total, 3B active	Mixture of experts	256K tokens	BF16, FP8	44 GB	None

JetBrains 2

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Mellum2-12B-A2.5B-InstructJetBrains/Mellum2-12B-A2.5B-Instruct	12B total, 2.5B active	Mixture of experts	128K tokens	BF16	29 GB	None
Mellum2-12B-A2.5B-ThinkingJetBrains/Mellum2-12B-A2.5B-Thinking	12B total, 2.5B active	Mixture of experts	128K tokens	BF16, FP8, NVFP4	8 GB	None

Jina AI 2

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Jina Embeddings v5 Text Smalljinaai/jina-embeddings-v5-text-small	0.7B	Dense	32K tokens	BF16	2 GB	None
Jina Reranker m0jinaai/jina-reranker-m0	2.4B	Dense	32K tokens	BF16	6 GB	None

Stability AI 2

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Stable Audio Openstabilityai/stable-audio-open-1.0	1.2B	Dense	Not stated	BF16	19 GB	None
Stable Diffusion 3.5stabilityai/stable-diffusion-3.5-medium	2.5B	Dense	Not stated	BF16	24 GB	None

StepFun 2

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Step-3.5-Flashstepfun-ai/Step-3.5-Flash	196B total, 11B active	Mixture of experts	256K tokens	BF16, FP8, INT4, INT8	118 GB	None
Step-3.7-Flashstepfun-ai/Step-3.7-Flash	198B total, 11B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	119 GB	None

Arcee AI 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Trinity-Large-Thinkingarcee-ai/Trinity-Large-Thinking	398B total, 13B active	Mixture of experts	256K tokens	BF16, NVFP4	239 GB	None

Boson AI 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Higgs Audio V3 TTSbosonai/higgs-audio-v3-tts-4b	4B	Dense	Not stated	BF16	24 GB	None

Dots 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
dots3-note-prevdots-studio/dots3-note-prev	288B total, 17B active	Mixture of experts	512K tokens	FP8, BF16	359 GB	None

Fish Audio 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Fish Speech S2 Profishaudio/s2-pro	4.6B	Dense	Not stated	BF16	48 GB	None

Hy (Tencent) 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Hy3tencent/Hy3	295B total, 21B active	Mixture of experts	256K tokens	BF16, FP8, NVFP4	214 GB	None

IBM Granite 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Granite-3.2-2B-Instructibm-granite/granite-3.2-2b-instruct	2B	Dense	128K tokens	BF16	6 GB	None

IndexTeam 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
IndexTTS-2.5IndexTeam/IndexTTS-2.5	Undisclosed	Dense	Not stated	BF16	96 GB	None

InternVL (OpenGVLab) 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
InternVL3.5OpenGVLab/InternVL3_5-8B	8B	Dense	40K tokens	BF16	19 GB	None

Lightricks 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
LTX-2.5Lightricks/LTX-2.5-Diffusers	19B	Dense	Not stated	BF16, FP8	61 GB	None

LongCat (Meituan) 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
LongCat-Image-Editmeituan-longcat/LongCat-Image-Edit	6B	Dense	Not stated	BF16	36 GB	None

MindLab Research 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Macaron-V1-Coding-Ventimindlab-research/Macaron-V1-Coding-Venti	743B total, 39B active	Mixture of experts	1M tokens	BF16, FP8	893 GB	None

Muse (Meta) 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Muse Glimmer 30Bmeta-models/Muse-Glimmer-30B	29.6B	Dense	128K tokens	BF16, FP8, NVFP4	31 GB	None

Preferred Networks 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
PLaMo 3 NICT 31B Basepfnet/plamo-3-nict-31b-base	32B	Dense	4K tokens	BF16	77 GB	None

Seed (ByteDance) 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Seed-OSS-36BByteDance-Seed/Seed-OSS-36B-Instruct	36B	Dense	512K tokens	BF16	86 GB	None

Wan (Alibaba) 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
Wan2.2Wan-AI/Wan2.2-T2V-A14B-Diffusers	28B total, 14B active	Mixture of experts	Not stated	BF16	152 GB	None

Z.ai 1

Model	Size	Architecture	Context	Precisions	Memory floor	Dynamo
GLM-4-9Bzai-org/glm-4-9b-hf	9B	Dense	8K tokens	BF16	22 GB	None

Questions

About this directory.

Where do these recipes come from?From the vLLM recipes project (vllm-project/recipes) and NVIDIA Dynamo's recipes directory, both Apache 2.0. The control plane ships a snapshot of each at a pinned commit; this directory is built from the same snapshot, captured 2026-09-01.
Are the memory figures measured on Wingsmith?No. Each memory floor is the minimum the vLLM recipe publishes for that precision, across the whole replica. The control plane shows it next to its own sizing and says so when the two disagree.
Can Wingsmith serve a model that is not listed here?Yes. Any HuggingFace slug can be registered. A recipe adds published flags and a memory floor to the draft when one exists; without one, the plane sizes the weights from their own config.

Recipe data from vllm-project/recipes and ai-dynamo/dynamo, both Apache 2.0.

Status

Open source,
under Apache 2.0.

Wingsmith is pre-release. The first public release, and everything after it, will be under Apache 2.0.

View on GitHub Read the docs Run it locally

Built and not built

Built

Multi-cluster placement and routing, Dynamo deployments, device inventory, endpoints and quotas, the console, the CLI, and three measurement tools.

Not built

Published images and charts, per-user roles and SSO, batch and training workloads, and a Helm chart for the control plane.

Pre-release. No tagged release, images, or charts yet, and APIs and protocols can change.

Serving recipes for178 open models.

Models by provider.

Qwen 32

GLM (Z-AI) 16

NVIDIA 12

LiquidAI 10

DeepSeek 9

Google 7

inclusionAI 7

Meta 7

Mistral AI 7

Moonshot AI 7

MiniMax 6

OpenMOSS 6

Hunyuan (Tencent) 4

Poolside 4

Ernie (Baidu) 3

Microsoft 3

MiMo (Xiaomi) 3

MiniCPM (OpenBMB) 3

OpenAI 3

PaddlePaddle 3

InternLM 2

JetBrains 2

Jina AI 2

Stability AI 2

StepFun 2

Arcee AI 1

Boson AI 1

Dots 1

Fish Audio 1

Hy (Tencent) 1

IBM Granite 1

IndexTeam 1

InternVL (OpenGVLab) 1

Lightricks 1

LongCat (Meituan) 1

MindLab Research 1

Muse (Meta) 1

Preferred Networks 1

Seed (ByteDance) 1

Wan (Alibaba) 1

Z.ai 1

Open source,under Apache 2.0.

Built and not built

Serving recipes for
178 open models.

Open source,
under Apache 2.0.