LocalFit
Submit a benchmark

granite-swash-2b Base

ibm-granite/granite-swash-2b

Released2026-07-07Size2.1B denseRoleBaseLicenseapache-2.0Context8K config max

Where it runs 8K context · ≥ 5 t/s floor

SetupBest planMemoryGenerate
8 GB GPU + 32 GB RAMnative BF16vLLMFull GPUFits≤ 66 t/s Ceiling
12 GB GPU + 32 GB RAMnative BF16vLLMFull GPUFits≤ 83 t/s Ceiling
16 GB GPU + 64 GB RAMnative BF16vLLMFull GPUFits≤ 206 t/s Ceiling
24 GB GPU + 64 GB RAMnative BF16vLLMFull GPUFits≤ 232 t/s Ceiling
32 GB GPU + 128 GB RAMnative BF16vLLMFull GPUFits≤ 412 t/s Ceiling
2× 24 GB GPUs + 128 GB RAMnative BF16vLLMFull GPUFits≤ 215 t/s Ceiling
96 GB workstation GPU + 256 GB RAMnative BF16vLLMFull GPUFits≤ 412 t/s Ceiling
Mac, 24 GBDoes not fit
Mac, 48 GBDoes not fit
Mac, 128 GBDoes not fit
Mac, 512 GBDoes not fit
DGX Spark 128 GBnative BF16vLLMUnified memoryFits≤ 63 t/s Ceiling
Ryzen AI Max+ 395, 128 GBnative BF16vLLMUnified memoryFits≤ 59 t/s Ceiling
CPU only, 64 GB DDR5-5600 dual channelDoes not fit

Same plan engine as the planner, with its default reserves and RAM. How plans are chosen.

Files exact Hugging Face bytes

ArtifactSizeFilesRevisionSource
Native safetensors BF16
ibm-granite/granite-swash-2b
3.99 GiB1af1e322710HF files

Not found in the representative quant search (not proven unavailable): GGUF FP16, GGUF Q8_0, GGUF Q5_K_M, GGUF Q4_K_M, MLX 8-bit, MLX 4-bit, EXL3, AWQ, GPTQ. Complete selected file sets; projectors, MTP/draft and lookup files are listed separately and budgeted only by plans that use them. File bytes are not allocated memory.

Quality

No Artificial Analysis evaluation mapped to this checkpoint.

Arena text rating

Unknown

Publisher-reported benchmarks

None extracted from the model card.

AA and Arena describe the hosted model in the stated mode, not each quantized file; scores are never inherited by base, pruned, abliterated or quantized variants and never averaged together.

How to run it

Transformers · unknown

publisher-documented, not executed source

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ibm-granite/granite-swash-2b",
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="flash_attention_3",  # or "flash_attention_4" or "eager", NOT "sdpa"
)
tokenizer = AutoTokenizer.from_pretrained("ibm-granite/granite-swash-2b")

inputs = tokenizer("The capital of France is", return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=50, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Prerequisites: transformers>5.8.0; suitable FlashAttention 3/4 kernel or CPU eager.

  • SDPA explicitly unsupported; use flash_attention_3, flash_attention_4 or eager.
  • Attention sinks and 7 full+17 window 128 layers must be reflected in KV.

Recipes are source examples (settings are the publisher’s or author’s, not measured optima) and were not executed by LocalFit. A downloadable file does not prove your backend build supports the architecture.

Reported speeds 0

No reported measurement yet. Submit one.