Sarvam 105B — Hardware Requirements & GPU Compatibility
Chat
Sarvam 105B is a 106.0B-parameter open language model from sarvamai. It supports a context window of up to 131,072 tokens. At BF16 it needs about 233.27 GB of VRAM — see which GPUs and Macs can run it below.
Sarvam 105B runs on a 256 GB+ Mac or a datacenter GPU
Best local pick → Mac Studio (M3 Ultra, 256GB) · BF16 · 238.6 GB · ~19 tok/s
Macs from 256 GB (BF16 · ~19 tok/s)
Apache 2.0
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Sarvam 105B
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve sarvamai/sarvam-105b --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 B300 (288 GB) (49 GB headroom) · BF16
- Generation speed
- ~143 tok/s
- generation speed
- Cost per 1M output tokens
- $13.52
- per 1M output tokens
How Much VRAM Does Sarvam 105B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 233.3 GB | 244.6 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on Instinct MI325X 256 GB |
Selected BF16
Brain floating point 16238.6GBat 16K context
- Weights
- 232.2 GB
- KV cache
- 6.1 GB
- Overhead
- 0.3 GB
Server-class: ~238.6 GB at 16K context · one Instinct MI325X 256 GB
Which Hardware Can Run Sarvam 105B?
BF16 · 233.3 GBSarvam 105B (BF16) requires 233.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 304+ GB is recommended. No consumer GPU has enough memory. 2 devices with unified memory can run Sarvam 105B, including Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 256GB).
Needs server-class hardware
Sarvam 105B runs on a 256 GB+ Mac or a datacenter GPU · BF16 needs ~238.6 GB
| Hardware | Quant | Memory | Speed | Context | Hardware page | |
|---|---|---|---|---|---|---|
| Mac 256 GB · Mac Studio (M3 Ultra, 256GB) | BF16 | 238.6 GB | ~19 tok/s | 32K ctx | Mac Studio → | BF16 · 238.6 GB · ~19 tok/s · 32K ctx |
| Instinct MI325X 256 GB · one GPU | BF16 | 238.6 GB | ~72 tok/s | 32K ctx | Instinct MI325X → | BF16 · 238.6 GB · ~72 tok/s · 32K ctx |
| 4× H100 SXM 80 GB · 320 GB | BF16 | 238.6 GB | — | 128K ctx | H100 SXM → | BF16 · 238.6 GB · 128K ctx |
| 2× H200 SXM 141 GB · 282 GB | BF16 | 238.6 GB | — | 32K ctx | H200 SXM → | BF16 · 238.6 GB · 32K ctx |
| 2× B200 192 GB · 384 GB | BF16 | 238.6 GB | — | 128K ctx | B200 → | BF16 · 238.6 GB · 128K ctx |
Also: Instinct MI325X · Instinct MI350X · Instinct MI355X · B300 · Mac Studio (M3 Ultra, 256GB) · Mac Studio (M3 Ultra, 512GB) · NVIDIA DGX A100 640GB · NVIDIA DGX H100
By VRAM tier · typical GPUs and Macs at BF16
Consumer GPUs (up to 32 GB) and Macs up to 192 GB: too small for this model (needs ≥ 238.6 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
Similar models
5| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Llama 4 Scout 17B 16E InstructMeta | 81.8 | 77 GB | —needs > 24 GB | same size |
| C4ai Command A 03 2025Cohere | 79.8 | 78.7 GB | —needs > 24 GB | same size |
| GLM 4.6VZ.ai | — | 66 GB | —needs > 24 GB | same size |
| C4ai Command R Plus 08 2024Cohere | — | 73.8 GB | —needs > 24 GB | same size |
| Qwen1.5 110BAlibaba | — | 72.4 GB | —needs > 24 GB | same size |
Specifications
- Publisher
- sarvamai
- Parameters
- 106.0B
- Active parameters
- 12.4B per token
- Architecture
- SarvamMLAForCausalLM
- Context Length
- 131,072 tokens
- Vocabulary Size
- 262,144
- Release Date
- 2026-03-03
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Sarvam 105B need?
Sarvam 105B requires 233.3 GB of VRAM at BF16.
VRAM = Weights + KV Cache + Overhead
Weights = 106.0B × 16 bits ÷ 8 = 212.1 GB
KV Cache + Overhead ≈ 21.2 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.
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run Sarvam 105B?
No — Sarvam 105B requires at least 233.3 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
Can I run Sarvam 105B on a Mac?
Yes — Mac Studio (M3 Ultra, 256GB) and 1 other Mac can run Sarvam 105B. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 233.3 GB of usable unified memory (RAM minus macOS overhead).
Can I run Sarvam 105B locally?
Yes — Sarvam 105B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Sarvam 105B?
At BF16, Sarvam 105B can reach ~83 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 B300 → 27.3 GB active ÷ (8000 × 0.65) = 5.25 ms, plus 32 layers × 0.055 ms = 1.76 ms, so 1000 ÷ 7.01 ms = ~143 tok/s
Estimated speed at BF16 (233.3 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Sarvam 105B?
At BF16, the download is about 212.06 GB.
Which GPUs can run Sarvam 105B?
No single consumer GPU has enough VRAM to run Sarvam 105B at BF16 (233.3 GB). Multi-GPU or professional hardware is required.
Which devices can run Sarvam 105B?
2 devices with unified memory can run Sarvam 105B at BF16 (233.3 GB), including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.