StableBeluga2 — Hardware Requirements & GPU Compatibility
Chat
StableBeluga2 is Stability AI's chat-tuned, 70-billion-parameter model, an Orca-style supervised fine-tune of Meta's Llama 2 70B base rather than a model trained from scratch. It follows Stability's internal reimplementation of the Orca training recipe, using a structured System/User/Assistant prompt format, and was trained in mixed BF16 precision with AdamW; smaller Stable Beluga 7B and 13B siblings were released alongside it. It is a historically significant early Llama-2 fine-tune from the 2023 open-model wave rather than a current state-of-the-art model by 2026 standards. At 70 billion parameters, it needs a multi-GPU workstation or server to run, even once quantized. Context length is 4,096 tokens, inherited from the Llama 2 base. It is released under Stability AI's Stable Beluga Non-Commercial Community License, which restricts the fine-tuned weights to non-commercial use. It was published in July 2023.
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run StableBeluga2
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/mradermacher/StableBeluga2-i1-GGUF:Q4_K_MQ4_K_M · needs ~43.6 GB · 4K context
Download: StableBeluga2.i1-Q4_K_M.gguf41.42 GBmradermacher
Ways to runOther ways
each opens step-by-stepRun in cloud
Your own private copy · Q4_K_M · 48 GB GPU
- The easiest way to try this model
- Running in a few minutes, easy setup
- Chat right here or connect any app
$5+ bonus for new users
Runs on H100 (80 GB) (36 GB headroom) · Q4_K_M
- Generation speed
- ~51 tok/s
- generation speed
- Cost per 1M output tokens
- $14.74
- per 1M output tokens
How Much VRAM Does StableBeluga2 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 30.7 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Does not fit: needs 30.7 GB | |
| 31.6 GB | — | 3-bit small — 3-bit small quantization | Does not fit: needs 31.6 GB | |
| 35.1 GB | — | 3-bit medium — 3-bit medium quantization | Does not fit: needs 35.1 GB | |
| 36.0 GB | — | 4-bit · legacy — 4-bit legacy quantization | Does not fit: needs 36 GB | |
| 43.0 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Does not fit: needs 43 GB | |
| 50.9 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 50.9 GB | |
| 58.7 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 58.7 GB | |
| 71.0 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 71 GB |
Selected Q4_K_M
4-bit medium quantization43.6GBat 4K context
- Weights
- 42.0 GB
- KV cache
- 1.3 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 19.6 GB more, IQ2_S fits
Or .
Which Hardware Can Run StableBeluga2?
Q4_K_M · 43.0 GBStableBeluga2 (Q4_K_M) requires 43.0 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 56+ GB is recommended. Using the full 4K context window can add up to 0.7 GB, bringing total usage to 43.6 GB. No consumer GPU has enough memory. 67 devices with unified memory can run StableBeluga2, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).
≤ 20 GB
≤ 20 GB tier →doesn't fit · smallest quant needs 20.9 GB
32–40 GB
Q4_K_M · best fit Q2_K (31.4 GB)
1 GPU · 2 Mac classes · 1 device family
RTX 5090~38Simulate output on RTX 5090M5 Max36 GB~11Simulate output on Mac Studio M5 Max (36 GB)M3 Pro*36 GB~3Simulate output on MacBook Pro 14" M3 Pro (36 GB)Orin~4Simulate output on NVIDIA Jetson AGX Orin 32GB48 GB
48 GB tier →fits · 4.4 GB free · up to 4K ctx
2 Mac classes · 1 device family
M5 Max48 GB~10Simulate output on Mac Studio M5 Max (48 GB)M5 Pro48 GB~5Simulate output on Mac Mini M5 Pro (48 GB)Snapdragon X2 Elite Extreme (X2E-96-100)~3Simulate output on Snapdragon X2 Elite Extreme Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
How good is StableBeluga2?
llmrun Score blends public benchmark boards into one number; 100 = the average open model.
Not enough data yet for an llmrun Score.
A Score needs results on enough independent boards. How the Composite Score is built
Benchmarks
The 6 boards that weigh most in the llmrun Score, of 7 with a result for this model; badges show who ran the test.
- MMLU
Epoch AI · Jul 2023Lab-reported
Lab-reported68.6%
#33/85 open
- BIG-Bench Hard
Epoch AI · Jul 2023Lab-reported
Lab-reported69.3%
#9/38 open
- GSM8K
Epoch AI · Jul 2023Lab-reported
Lab-reported69.6%
#19/62 open
- HellaSwag
Epoch AI · Jul 2023Lab-reported
Lab-reported84.1%
#9/47 open
- ARC Challenge
Epoch AI · Jul 2023Lab-reported
Lab-reported86.1%
#8/56 open
- BoolQ
Epoch AI · Jul 2023Lab-reported
Lab-reported89.4%
#1/38 open
Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Llama 3.3 70B InstructMeta | 80.0 | 51.3 GB | —needs > 24 GB | same size |
| NVIDIA Nemotron 3.5 Bolt 74B A9B BF16NVIDIA | — | 54.2 GB | —needs > 24 GB | same size |
| Qwen2.5 VL 72B InstructAlibaba | — | 49.7 GB | —needs > 24 GB | same size |
| Qwen2 VL 72B InstructAlibaba | — | 49.7 GB | —needs > 24 GB | same size |
| Qwen1.5 72B ChatAlibaba | — | 86.6 GB | —needs > 24 GB | same size |
| CodeLlama 70B Instruct HFMeta | — | 23.2 GB | ~29 | same size |
Specifications
- Publisher
- Stability AI
- Parameters
- 70B
- Architecture
- LlamaForCausalLM
- Context Length
- 4,096 tokens
- Vocabulary Size
- 32,000
- Release Date
- 2023-07-20
Frequently Asked Questions
How much VRAM does StableBeluga2 need?
StableBeluga2 requires 43.0 GB of VRAM at Q4_K_M, or 141.0 GB at BF16. Full 4K context adds up to 0.7 GB (43.6 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 70B × 4.8 bits ÷ 8 = 42 GB
KV Cache + Overhead ≈ 1 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 ≈ 1.6 GB (at full 4K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run StableBeluga2?
Yes, at IQ2_S (22.9 GB) or lower. Higher quantizations like IQ2_M (24.6 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for StableBeluga2?
For StableBeluga2, Q4_K_M (43.0 GB) offers the best balance of quality and VRAM usage. Q5_K_S (49.1 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 20.2 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run StableBeluga2 on a Mac?
Yes — Mac Studio M1 Max (32 GB) and 84 other Macs can run StableBeluga2, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 43.0 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M5 Max (48 GB).
Can I run StableBeluga2 locally?
Yes — StableBeluga2 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is StableBeluga2?
At Q4_K_M, StableBeluga2 can reach ~112 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 43.0 × 0.65 = ~121 tok/s
Estimated speed at Q4_K_M (43.0 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of StableBeluga2?
At Q4_K_M, the download is about 42.00 GB. The full-precision BF16 version is 140.00 GB. The smallest option (IQ2_XXS) is 19.25 GB.
Which GPUs can run StableBeluga2?
No single consumer GPU has enough VRAM to run StableBeluga2 at Q4_K_M (43.0 GB). Multi-GPU or professional hardware is required.
Which devices can run StableBeluga2?
67 devices with unified memory can run StableBeluga2 at Q4_K_M (43.0 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.