Grug 35B — Hardware Requirements & GPU Compatibility
ChatCodeFunctionsReasoning
Grug 35B is a 35.1B-parameter open language model from ProCreations. It supports a context window of up to 262,144 tokens. At BF16 it needs about 70.60 GB of VRAM — see which GPUs and Macs can run it below.
Grug 35B runs on a 96 GB+ Mac or a workstation GPU
Best local pick → Mac Studio (M3 Ultra, 96GB) · BF16 · 70.7 GB · ~46 tok/s
Macs from 96 GB (BF16 · ~46 tok/s)
MIT
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Grug 35B
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve ProCreations/grug-35b --max-model-len 16384 --tensor-parallel-size 8vLLM repo with its own quantisation
Download: Weights on Hugging Face
Other ways
each opens step-by-stepRun in cloud
Fits on RTX PRO 6000 (96 GB) (25 GB headroom) · BF16
- Generation speed
- ~112 tok/s
- generation speed
- Cost per 1M output tokens
- $3.32
- per 1M output tokens
How Much VRAM Does Grug 35B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 70.6 GB | 70.9 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on RTX PRO 5000 Blackwell 72 GB |
Selected BF16
Brain floating point 1670.7GBat 16K context
- Weights
- 70.2 GB
- KV cache
- 0.2 GB
- Overhead
- 0.3 GB
Server-class: ~70.7 GB at 16K context · one RTX PRO 5000 Blackwell 72 GB
Which Hardware Can Run Grug 35B?
BF16 · 70.6 GBGrug 35B (BF16) requires 70.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 92+ GB is recommended. Using the full 262K context window can add up to 2.6 GB, bringing total usage to 73.2 GB. No consumer GPU has enough memory. 30 devices with unified memory can run Grug 35B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 96GB).
96 GB
fits · 25.3 GB free · up to 256K ctx
3 GPUs · 2 Mac classes
RTX PRO 6000 Blackwell Max-Q Workstation Edition~112Simulate output on RTX PRO 6000 Blackwell Max-Q Workstation EditionRTX PRO 6000 Blackwell Workstation Edition~112Simulate output on RTX PRO 6000 Blackwell Workstation EditionRTX PRO 6000 Blackwell Server Edition~103Simulate output on RTX PRO 6000 Blackwell Server EditionM3 Ultra*96 GB~46Simulate output on Mac Studio (M3 Ultra, 96GB)M3 Max*96 GB~22Simulate output on MacBook Pro 14" M3 Max (96 GB)128–144 GB
fits · 57.3 GB free · up to 256K ctx
4 GPUs · 2 Mac classes · 4 device families
GH200 Grace Hopper Superchip~214Simulate output on GH200 Grace Hopper SuperchipH200 NVL~212Simulate output on H200 NVLH200 SXM~212Simulate output on H200 SXMM2 Ultra*128 GB~45Simulate output on Mac Pro M2 Ultra (128 GB)M5 Max128 GB~38Simulate output on Mac Studio M5 Max (128 GB)GB10~22Simulate output on NVIDIA DGX SparkJetson T5000 (Thor)~22Simulate output on NVIDIA Jetson AGX Thor Developer Kit
By VRAM tier · typical GPUs and Macs at BF16
Consumer GPUs (up to 32 GB) and Macs up to 64 GB: too small for this model (needs ≥ 70.7 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Qwen3.6 35B A3BAlibaba | 106.8 | 22 GB | ~151 | same size |
| MiniCPM V 4.7 35B A3BOpenBMB | — | 22.1 GB | ~169 | same size |
| Ornith 1.5 35B A3Bornith-ai | — | 22 GB | ~151 | same size |
| Agents A1InternScience | — | 21.5 GB | ~171 | same size |
| Intern S2 PreviewInternLM | — | 73.2 GB | —needs > 24 GB | same size |
| EXAONE 4.5 33BLGAI-EXAONE | — | 23.9 GB | ~33 | same size |
Specifications
- Publisher
- ProCreations
- Parameters
- 35.1B
- Active parameters
- 3.9B per token
- Architecture
- Qwen3_5MoeForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 248,320
- Release Date
- 2026-07-15
- License
- MIT
Frequently Asked Questions
How much VRAM does Grug 35B need?
Grug 35B requires 70.6 GB of VRAM at BF16. Full 262K context adds up to 2.6 GB (73.2 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 35.1B × 16 bits ÷ 8 = 70.2 GB
KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 3 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run Grug 35B?
No — Grug 35B requires at least 70.6 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
Can I run Grug 35B on a Mac?
Yes — Mac Studio (M3 Ultra, 96GB) and 20 other Macs can run Grug 35B. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 70.6 GB of usable unified memory (RAM minus macOS overhead).
Can I run Grug 35B locally?
Yes — Grug 35B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Grug 35B?
At BF16, Grug 35B can reach ~104 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 7.8 GB active ÷ (8000 × 0.65) = 1.51 ms, plus 40 layers × 0.055 ms = 2.20 ms, so 1000 ÷ 3.71 ms = ~270 tok/s
Estimated speed at BF16 (70.6 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Grug 35B?
At BF16, the download is about 70.21 GB.
Which GPUs can run Grug 35B?
No single consumer GPU has enough VRAM to run Grug 35B at BF16 (70.6 GB). Multi-GPU or professional hardware is required.
Which devices can run Grug 35B?
30 devices with unified memory can run Grug 35B at BF16 (70.6 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.