Hy3-preview has 298.8B parameters (20B active per token). In BF16 its weights alone take 550 GiB; quantized to 4-bit, about 166 GiB. Each 32K-token request adds 10.0 GiB of KV cache. Cheapest way to run it today: 8 × GeForce RTX 3090 (24 GB) at $0.872 per hour. It calls tools and lets you set its reasoning effort.
Weights plus the KV cache of one request (BF16 cache), in GiB. Add 1 to 3 GiB for the inference engine itself.
Weights format
Weights only
+ 8K context
+ 32K context
+ 128K context
+ 256K context
BF16 / FP16
550
552
560
590
630
FP8 / INT8
276
278
286
316
356
4-bit (GGUF Q4_K_M)
166
168
176
206
246
How many GPUs to run Hy3-preview
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
GPU
BF16 / FP16
8-bit
4-bit
GeForce RTX 4090 24 GB
32 (several machines)
16 (several machines)
8
GeForce RTX 5090 32 GB
24 (several machines)
16 (several machines)
6
RTX PRO 6000 Blackwell 96 GB
6
3
2
Mac M5 Max 128 GB 96 GB usable
6
3
2
Ryzen AI Max+ 395 128 GB 96 GB usable
6
3
2
DGX Spark (GB10) 128 GB
5
3
2
NVIDIA H100 80 GB
16 (several servers)
5
3
NVIDIA H200 141 GB
5
3
2
NVIDIA B200 (HGX) 180 GB
4
2
2
8-bit: FP8 with vLLM (INT8 on GPUs without FP8), Q8_0 with llama.cpp. 4-bit: AWQ / GPTQ with vLLM, Q4_K_M with llama.cpp, both counted at the Q4_K_M size.
Cheapest way to run Hy3-preview today
For each GPU: the fewest cards that fit Hy3-preview with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-11-22).
On-demand prices from AWS, Google Cloud, Azure, RunPod, Vast.ai, Oracle Cloud, OVHcloud, Scaleway, Verda, Akash and UpCloud, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
vLLM and SGLang apply these defaults unless a request sets its own. With llama.cpp, pass them yourself: --temp 0.9 --top-p 1
Reasoning. Adjustable reasoning effort. Reasoning tokens take context and KV cache like any other token: leave room for them.
KV cache: how memory grows with context
Architecture: 80 layers · 80 full attention (8 KV × 128). Every layer keeps keys and values for every token (grouped-query attention), so the cache grows linearly with context. Only 20B parameters are active per token, but all 299B must sit in memory: VRAM depends on the total, speed on the active part.
Context per request
BF16 cache
FP8 cache
8K tokens
2.50 GiB
1.25 GiB
32K tokens
10.0 GiB
5.00 GiB
128K tokens
40.0 GiB
20.0 GiB
256K tokens
80.0 GiB
40.0 GiB
Hy3-preview VRAM FAQ
How much VRAM does Hy3-preview need?
In BF16 the weights alone take 550 GiB (590 GB). In FP8 that is 276 GiB, and about 166 GiB with 4-bit quantization (Q4_K_M). Each request then adds KV cache: 2.50 GiB at 8K tokens and 10.0 GiB at 32K (BF16 cache).
Can Hy3-preview run on a single RTX 4090 (24 GB)?
No. Even in 4-bit it needs 8 RTX 4090s. In 4-bit it needs 3 H100 GPUs.
Can Hy3-preview run on an RTX 5090 (32 GB)?
No. Even in 4-bit it needs 6 RTX 5090s.
Can Hy3-preview run on a Mac?
Yes. In 4-bit (Q4_K_M with llama.cpp) it needs a Mac with 256 GB of unified memory or more, of which macOS gives the GPU about 192 GB, for one request with an 8K-token context. In 8-bit (Q8_0) it needs 512 GB.
Can Hy3-preview run on a DGX Spark or a Ryzen AI Max+ 395 (128 GB)?
No. Even in 4-bit it needs 2 DGX Spark units or 2 Ryzen AI Max+ 395 machines (128 GB), linked over the network.
What is the cheapest way to run Hy3-preview?
On 2026-11-22, the cheapest on-demand setup is 8 × GeForce RTX 3090 (24 GB) with Q4_K_M weights on Vast.ai, at $0.872 per hour (about $636.56 per month). Next: 8 × RTX A5000 (24 GB) with Q4_K_M weights on Vast.ai, at $1.51 per hour. Sized for one 8K-token request; prices are checked every hour.
How many H100 GPUs does Hy3-preview need?
With one request and an 8K-token context: 16 in BF16, 5 in FP8 and 3 in 4-bit (vLLM, 90% of memory usable). Above 8 GPUs, that means several servers. Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
Does Hy3-preview support tool calling?
Yes, Hy3-preview supports tool (function) calling: it can ask your application to run a function and use the result.
Does Hy3-preview have a thinking mode?
Adjustable reasoning effort. Reasoning tokens count toward the context and the KV cache.
What temperature should I use with Hy3-preview?
The official defaults are temperature 0.9, top-p 1.
Hy3-preview VRAM badge
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/hy3-preview.json.
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.