Tencent, China · Released August 27, 2026
Hy4 Preview has 780.0B parameters (49B active per token). In BF16 its weights alone take 1434 GiB; quantized to 4-bit, about 433 GiB. Each 32K-token request adds 2.83 GiB of KV cache. Cheapest way to run it today: 7 × NVIDIA A100 80GB at $4.41 per hour. It calls tools and lets you set its reasoning effort.
Size your own setup in the calculator →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 | + 1M context |
|---|---|---|---|---|---|
| FP8 (native) | 748 | 749 | 751 | 759 | 839 |
| BF16 / FP16 | 1434 | 1435 | 1437 | 1445 | 1525 |
| FP8 / INT8 | 718 | 719 | 721 | 730 | 809 |
| 4-bit (GGUF Q4_K_M) | 433 | 434 | 436 | 444 | 524 |
Exact size of each quantization in 6block/Hy4-preview-GGUF (opens in a new tab), and the smallest GPU that runs it with llama.cpp, Ollama or LM Studio and an 8K-token context.
| Quantization | File size | Smallest GPU |
|---|---|---|
| IQ4_XS | 388 GiB | More than 96 GB: see the calculator |
| Q2_K | 270 GiB | More than 96 GB: see the calculator |
| IQ2_XS | 219 GiB | More than 96 GB: see the calculator |
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
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 | FP8 (native) | BF16 / FP16 | 8-bit | 4-bit |
|---|---|---|---|---|
| GeForce RTX 4090 24 GB | Not on this GPU | 96 (several machines) | 48 (several machines) | 24 (several machines) |
| GeForce RTX 5090 32 GB | Not on this GPU | 64 (several machines) | 32 (several machines) | 16 (several machines) |
| RTX PRO 6000 Blackwell 96 GB | Not on this GPU | 16 (several machines) | 8 | 5 |
| Mac M5 Max 128 GB 96 GB usable | Not on this GPU | 16 (several machines) | 8 | 5 |
| Ryzen AI Max+ 395 128 GB 96 GB usable | Not on this GPU | 16 (several machines) | 8 | 5 |
| DGX Spark (GB10) 128 GB | Not on this GPU | 16 (several machines) | 6 | 4 |
| NVIDIA H100 80 GB | 16 (several servers) | 24 (several servers) | 16 (several servers) | 7 |
| NVIDIA H200 141 GB | 7 | 16 (several servers) | 6 | 4 |
| NVIDIA B200 (HGX) 180 GB | 5 | 16 (several servers) | 5 | 3 |
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. Not on this GPU: the official FP8 weights do not run there (llama.cpp needs a GGUF version, and vLLM needs a GPU with FP8 support).
For each GPU: the fewest cards that fit Hy4 Preview with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-10-30).
| Setup | Per hour | Weights | Provider | Per month | Rent |
|---|---|---|---|---|---|
| 7 × NVIDIA A100 80GB | $4.41 | 4-bit | Vast.ai | $3,219.30 | Rent 7 × NVIDIA A100 80GB on Vast.ai (opens in a new tab) → |
| 5 × RTX PRO 6000 Blackwell 96 GB | $6.67 | Q4_K_M | Vast.ai | $4,869.10 | Rent 5 × RTX PRO 6000 Blackwell on Vast.ai (opens in a new tab) → |
| 3 × AMD MI300X 192 GB | $7.17 | 4-bit | RunPod | $5,234.10 | Rent 3 × AMD MI300X on RunPod (opens in a new tab) → |
| 7 × NVIDIA H100 80 GB | $14.95 | 4-bit | Vast.ai | $10,914.96 | Rent 7 × NVIDIA H100 on Vast.ai (opens in a new tab) → |
| 6 × NVIDIA H100 NVL 94 GB | $15.82 | 4-bit | Vast.ai | $11,550.06 | Rent 6 × NVIDIA H100 NVL on Vast.ai (opens in a new tab) → |
| 2 × NVIDIA B300 / GB300 288 GB | $16.70 | 4-bit | Scaleway | $12,191.00 | Rent 2 × NVIDIA B300 / GB300 on Scaleway (opens in a new tab) → |
| 2 × AMD MI355X / MI350X 288 GB | $17.20 | 4-bit | Oracle Cloud | $12,556.00 | Rent 2 × AMD MI355X / MI350X on Oracle Cloud (opens in a new tab) → |
| 4 × NVIDIA H200 141 GB | $20.22 | 4-bit | Vast.ai | $14,757.68 | Rent 4 × NVIDIA H200 on Vast.ai (opens in a new tab) → |
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.
AWS, Google Cloud and UpCloud prices: Spare Cores (opens in a new tab), CC BY-SA 4.0 (opens in a new tab), shared under the same license.
No GPU to rent: these providers serve Hy4 Preview and bill per token. Speeds are measured by Hugging Face for one request.
| Provider | Output, per 1M tokens | Input, per 1M | Speed | First token |
|---|---|---|---|---|
| Novita | $2.50 | $0.834 | 26 tokens/s | 2.1 s |
| DeepInfra | $2.50 | $0.834 | 35 tokens/s | 391 ms |
Prices and speeds from Hugging Face Inference Providers (opens in a new tab), read every day (October 30, 2026). An API wins at low volume; renting GPUs wins when they stay busy: the calculator gives the break-even.
Sampling defaults from the official generation_config.json (opens in a new tab) of tencent/Hy4-preview.
| Setting | Value |
|---|---|
| Temperature | 0.9 |
| Top-p | 1 |
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.
Architecture: MoE 256 experts + 1 shared · gated DSA: MLA (512 + 64 latent) on 78 layers · sparse-attention indexer keys on 21. Its cache stays compact because MLA stores one compressed latent per token and layer instead of full keys and values. Only 49B parameters are active per token, but all 780B must sit in memory: VRAM depends on the total, speed on the active part.
| Context per request | BF16 cache | FP8 cache |
|---|---|---|
| 8K tokens | 0.71 GiB | 0.41 GiB |
| 32K tokens | 2.83 GiB | 1.65 GiB |
| 128K tokens | 11.3 GiB | 6.58 GiB |
| 1M tokens | 90.5 GiB | 52.7 GiB |
In BF16 the weights alone take 1434 GiB (1540 GB). In FP8 that is 718 GiB, and about 433 GiB with 4-bit quantization (Q4_K_M); the official FP8 checkpoint is 748 GiB. Each request then adds KV cache: 0.71 GiB at 8K tokens and 2.83 GiB at 32K (BF16 cache).
No. Even in 4-bit it needs 24 RTX 4090s, spread over several machines. In 4-bit it needs 7 H100 GPUs.
No. Even in 4-bit it needs 16 RTX 5090s, spread over several machines.
No single Mac holds it, even in 4-bit: it needs 434 GiB with an 8K-token context, and the largest Mac (512 GB) gives its GPU about 384 GB.
No. Even in 4-bit it needs 4 DGX Spark units or 5 Ryzen AI Max+ 395 machines (128 GB), linked over the network.
On 2026-10-30, the cheapest on-demand setup is 7 × NVIDIA A100 80GB with 4-bit weights on Vast.ai, at $4.41 per hour (about $3,219.30 per month). Next: 5 × RTX PRO 6000 Blackwell (96 GB) with Q4_K_M weights on Vast.ai, at $6.67 per hour. Sized for one 8K-token request; prices are checked every hour.
With one request and an 8K-token context: 24 in BF16, 16 in FP8 and 7 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.
Yes, Hy4 Preview supports tool (function) calling: it can ask your application to run a function and use the result.
Adjustable reasoning effort. Reasoning tokens count toward the context and the KV cache.
The official defaults are temperature 0.9, top-p 1.
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/hy4-preview.json.
Tencent, China · Released August 27, 2026[](https://studiotvai.com/vram-requirements/hy4-preview)Same family (large moe): MiniMax M3.1 Flash · MiniMax M3 · MiniMax M2.7 · DeepSeek R1 · Kimi K2.6 · GLM-5.3 · Atria Dawn · Atria Dawn Preview · K-EXAONE 2.0 750B-A37B · Inkling · GLM-5.3-Flash · DeepSeek V4 Flash · DeepSeek V4 Pro · DeepSeek V4.1 Flash · MiMo-V2.6 Pro · MiMo-V2.6 Flash · Muse Spark Open · Beam 1.0 501B-A23B · Kimi K3 · Mistral Large 4 · Mistral Large 3 · Mistral Small 4 119B · Leanstral 1.5 119B-A6B · Mixtral 8x22B · DeepSeek V4 Flash Vision Exp · Qwen3.8 2.4T-A95B · Ornith 1.5 397B · Intern S2 397B · DeepSeek V3 · DeepSeek V3.1 · Kimi K2.7 Code · Kimi K2.5 · Kimi K2 · GLM-5.2 · GLM-5.1 · GLM-5 · MiniMax M2.5 · MiniMax M2 · MiMo-V2.5 · Ornith-1.0 397B · Qwen3 235B-A22B · Qwen3.5 397B-A17B · GLM-4.7 · GLM-4.5 · Ling-3.1-flash · Step 5 Preview · Hy3
Llama 3.1 8B · Llama 3 8B · Llama 3.2 3B · Nanbeige4.3 33B · Granite 4.2 30B · Granite 4.2 8B · Granite 4.2 3B · MiniCPM5 2B · MiniCPM5 1B · Ministral 3 14B · Ministral 3 8B · Ministral 3 3B · Devstral 2 123B · Devstral Small 2 24B · Mistral Medium 3.5 128B · Magistral Small 1.2 24B · Mistral Small 3 24B · Mistral Large 2411 123B · Pixtral Large 124B · Codestral 22B · Mistral Small 2409 22B · Pixtral 12B · Nemotron 3.5 Parse · Gemma 2 9B · Gemma 3 4B · LLaVA 1.6 Mistral 7B · Mistral Small 3.2 24B · Qwen3 32B · Llama 3.3 70B · Llama 3.1 405B · Hy-Memory 30B-A3B · Kimi VL A3B Instruct · Mixtral 8x7B · Unlimited-OCR · Unlimited-OCR 2 · Xing4.0 29B-A4B · Qwen3.5 4B · MiniCPM-V-4.7 1B · Raon-1 mini preview · Qwen3.5 9B · Qwen3.8 27B · Qwen3.6 35B-A3B · Qwen3.5 122B-A10B · Qwen3.8-Flash-Next · LFM3 1.7B · LFM3-VL 2B · d1 2B · LFM2.5 8B-A1B · LFM2 24B-A2B · Gemma 4 12B · Gemma 4 31B · Gemma 4 26B-A4B · gpt-oss 20B · gpt-oss 120B · Llama 4 Scout 17B-16E · Llama 4 Maverick 17B-128E · Kolibri-1 · Recursive Flint · Nemotron Labs Dory 1.0 4B · Step 3.5 Flash · Step 3.7 Flash · Laguna S 2.1 · Laguna XS 2.1 · Command A+ · Inkling Small · Muse Glimmer 30B · Nemotron 3 Nano 30B-A3B · Nemotron 3 Super 120B-A12B · Nemotron 3 Ultra 550B-A55B · Mamba-Codestral 7B · Nemotron 3.5 Super VL 120B-A12B · Nemotron 3.5 Bolt 74B-A9B · MiniMax M3.1 Flash · MiniMax M3 · MiniMax M2.7 · DeepSeek R1 · Kimi K2.6 · GLM-5.3 · Atria Dawn · Atria Dawn Preview · K-EXAONE 2.0 750B-A37B · Inkling · GLM-5.3-Flash · DeepSeek V4 Flash · DeepSeek V4 Pro · DeepSeek V4.1 Flash · MiMo-V2.6 Pro · MiMo-V2.6 Flash · Muse Spark Open · Beam 1.0 501B-A23B · Kimi K3 · Mistral Large 4 · Mistral Large 3 · Mistral Small 4 119B · Leanstral 1.5 119B-A6B · Mixtral 8x22B · Qwen3-Coder-Next 80B-A3B · Qwen3-Coder 30B-A3B · Qwen3-Coder 480B-A35B · DeepSeek V4 Flash Vision Exp · Qwen3.8 2.4T-A95B · Ornith 1.5 397B · Ornith 1.5 35B-A3B · Ornith 1.5 9B · Intern S2 397B · Clef · Clef Flash · d1 3B · Humanizer · JEV 27B VL · Mellum2.1 12B-A2.5B · LightOnOCR-3 4B · Spark-X2.5 4B · Agnes-3.0-Qwen · Youtu-Parsing-Omni · Agnes-3.0-Flash · Mistral 7B v0.3 · Mistral Small 3.1 24B · Mistral Nemo · Ministral 8B · DeepSeek V3 · DeepSeek R1 Qwen3 8B · DeepSeek R1 Distill Qwen 32B · DeepSeek V3.1 · DeepSeek R1 Distill Qwen 14B · DeepSeek R1 Distill Qwen 7B · DeepSeek R1 Distill Llama 8B · Qwen3 4B · Qwen3 8B · Qwen2.5 7B · Qwen3 14B · Qwen2.5 3B · Qwen3.6 27B · Qwen3.5 27B · Qwen2.5-Coder 7B · Qwen2.5 32B · Qwen2.5-Coder 14B · Qwen3.5 35B-A3B · Qwen3 30B-A3B · Qwen2.5 14B · Qwen2.5-Coder 32B · Kimi K2.7 Code · Kimi K2.5 · Kimi K2 · GLM-5.2 · GLM-5.1 · GLM-5 · MiniMax M2.5 · MiniMax M2 · MiMo-V2.5 · Ornith-1.0 397B · Qwen2.5-Coder 3B · Qwen3 235B-A22B · Qwen3.5 397B-A17B · Gemma 4 E4B · Gemma 4 E2B · Nemotron 3 Nano 4B · Nemotron 3.5 Lightning 30B-A3B · MiniCPM-V-4.7 35B-A3B · LFM2.5 24B-A2B · GLM-4.7 · Qwen3-VL 8B · Qwen2.5-VL 7B · Qwen3-VL 4B · Qwen2.5-VL 3B · Qwen2.5-VL 32B · Qwen2-VL 7B · GLM-4.5-Air · GLM-4.5 · phi-4 · Phi-3.5-mini · Phi-4-mini · Phi-3-mini-128k · Ling-3.1-flash · Step 5 Preview · Hy3 · Ornith-1.5 122B · Spark-X2.6 4B · Sage-A1.5 · MLExpert · Agents-A1 4B · Ling-3.0-flash-Sante · MiniCPM-OCR-1.0 · WeVisOne-VL 4B · Granite 4.1 3B
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.