Intern S2 Preview — Hardware Requirements & GPU Compatibility
VisionFunctions
Intern-S2-Preview is InternLM's 36.1-billion-parameter scientific multimodal foundation model, a mixture-of-experts model continued-pretrained from Qwen3.5 (256 experts, 8 active per token, about 4.9 billion active parameters) and further trained through a full pipeline from pretraining to reinforcement learning on hundreds of professional scientific tasks. InternLM reports it matches its own trillion-scale Intern-S1-Pro on several core scientific benchmarks despite its much smaller size, adds material crystal-structure generation and small-molecule spatial modeling, and improves agentic tool use for scientific workflows; it also models long, heterogeneous time-series signals. At its size it needs a high-end consumer GPU or multi-GPU setup once quantized. Context length is 262,144 tokens, though the card evaluates text-reasoning benchmarks at up to 128,000 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in May 2026.
Intern S2 Preview runs on a 96 GB+ Mac or a workstation GPU
Best local pick → Mac Studio (M3 Ultra, 96GB) · BF16 · 73.2 GB · ~40 tok/s
Macs from 96 GB (BF16 · ~40 tok/s)
Apache 2.0
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Intern S2 Preview
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve internlm/Intern-S2-Preview --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) (22 GB headroom) · BF16
- Generation speed
- ~94 tok/s
- generation speed
- Cost per 1M output tokens
- $3.71
- per 1M output tokens
How Much VRAM Does Intern S2 Preview Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 72.6 GB | 73.8 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on H100 SXM 80 GB |
Selected BF16
Brain floating point 1673.2GBat 16K context
- Weights
- 72.2 GB
- KV cache
- 0.7 GB
- Overhead
- 0.3 GB
Server-class: ~73.2 GB at 16K context · one H100 SXM 80 GB
Which Hardware Can Run Intern S2 Preview?
BF16 · 72.6 GBIntern S2 Preview (BF16) requires 72.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 95+ GB is recommended. Using the full 262K context window can add up to 10.7 GB, bringing total usage to 83.2 GB. No consumer GPU has enough memory. 30 devices with unified memory can run Intern S2 Preview, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 96GB).
96 GB
fits · 22.8 GB free · up to 256K ctx
3 GPUs · 2 Mac classes
RTX PRO 6000 Blackwell Max-Q Workstation Edition~94Simulate output on RTX PRO 6000 Blackwell Max-Q Workstation EditionRTX PRO 6000 Blackwell Workstation Edition~94Simulate output on RTX PRO 6000 Blackwell Workstation EditionRTX PRO 6000 Blackwell Server Edition~86Simulate output on RTX PRO 6000 Blackwell Server EditionM3 Ultra*96 GB~40Simulate output on Mac Studio (M3 Ultra, 96GB)M3 Max*96 GB~18Simulate output on MacBook Pro 14" M3 Max (96 GB)128–144 GB
fits · 54.8 GB free · up to 256K ctx
4 GPUs · 2 Mac classes · 4 device families
GH200 Grace Hopper Superchip~189Simulate output on GH200 Grace Hopper SuperchipH200 NVL~187Simulate output on H200 NVLH200 SXM~187Simulate output on H200 SXMM2 Ultra*128 GB~39Simulate output on Mac Pro M2 Ultra (128 GB)M5 Max128 GB~32Simulate output on Mac Studio M5 Max (128 GB)GB10~17Simulate output on NVIDIA DGX SparkJetson T5000 (Thor)~17Simulate 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 ≥ 73.2 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
Alternatives to Intern S2 Preview
Intern-S2 family · 3 models →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 |
| EXAONE 4.5 33BLGAI-EXAONE | — | 23.9 GB | ~33 | same size |
| Qwen3.5 35B A3B BaseAlibaba | — | 22 GB | ~151 | same size |
Specifications
- Publisher
- InternLM
- Family
- Intern-S2
- Parameters
- 36.1B
- Active parameters
- 4.9B per token
- Architecture
- InternS2PreviewForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 251,392
- Release Date
- 2026-05-15
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Intern S2 Preview need?
Intern S2 Preview requires 72.6 GB of VRAM at BF16. Full 262K context adds up to 10.7 GB (83.2 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 36.1B × 16 bits ÷ 8 = 72.2 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 ≈ 11 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run Intern S2 Preview?
No — Intern S2 Preview requires at least 72.6 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
Can I run Intern S2 Preview on a Mac?
Yes — Mac Studio (M3 Ultra, 96GB) and 20 other Macs can run Intern S2 Preview. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 72.6 GB of usable unified memory (RAM minus macOS overhead).
Can I run Intern S2 Preview locally?
Yes — Intern S2 Preview can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Intern S2 Preview?
At BF16, Intern S2 Preview can reach ~100 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 → 9.8 GB active ÷ (8000 × 0.65) = 1.89 ms, plus 40 layers × 0.055 ms = 2.20 ms, so 1000 ÷ 4.09 ms = ~244 tok/s
Estimated speed at BF16 (72.6 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Intern S2 Preview?
At BF16, the download is about 72.20 GB.
Which GPUs can run Intern S2 Preview?
No single consumer GPU has enough VRAM to run Intern S2 Preview at BF16 (72.6 GB). Multi-GPU or professional hardware is required.
Which devices can run Intern S2 Preview?
30 devices with unified memory can run Intern S2 Preview at BF16 (72.6 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.