TO

tokyotech llm Swallow 70b instruct v0.1

TokyotechGGUF by mmngaUpdated 2024-05-04
Download on Hugging Face
Parameters69.2B
Context4K
Layers80
Architecturellama
Smallest quant14.6 GB
Quick answer

tokyotech llm Swallow 70b instruct v0.1 needs about 44.7 GB of memory at Q4_K_M (a 41.5 GB file) with an 8K context window, so it fits on a 48 GB GPU. This repository has 20 quants, from 14.6 GB (IQ1_S) to 48.9 GB (Q5_K_M).

Tokyotech Swallow 70b Instruct v0.1 – 69.2B parameter Japanese‑centric LLM

The Swallow 70b Instruct v0.1 is a 69.2B parameter llama‑based model trained on 4096 token context. It is released under the llama2 license and supports English and Japanese. The model is available in multiple GGUF quantizations.

Good at

  • Japanese instruction following
  • Large context handling
  • High‑capacity reasoning

Limits

  • Requires large VRAM for full precision
  • No vision support
  • License restricts commercial use

Best for

  • Japanese content generation
  • Long‑form dialogue
  • Research prototyping

Who it’s for. Developers needing a large Japanese‑focused LLM for research or internal tools. The model is suitable for those who can allocate sufficient GPU memory and comply with the llama2 license.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ1_S14.6 GB 16.5 GB
Too big Get file →
IQ1_M16.0 GB 17.9 GB
Too big Get file →
IQ2_XXS18.4 GB 20.2 GB
Too big Get file →
IQ2_XS20.4 GB 22.3 GB
Too big Get file →
IQ2_S21.5 GB 23.3 GB
Too big Get file →
IQ2_M23.3 GB 25.2 GB
Too big Get file →
Q2_K25.6 GB 27.4 GB
Too big Get file →
IQ3_XXS26.7 GB 28.5 GB
Too big Get file →
IQ3_XS28.4 GB 30.3 GB
Too big Get file →
IQ3_S30.0 GB 31.9 GB
Too big Get file →
Q3_K_S30.0 GB 31.9 GB
Too big Get file →
IQ3_M31.1 GB 32.9 GB
Too big Get file →
Q3_K_M33.4 GB 35.2 GB
Too big Get file →
Q3_K_L36.3 GB 38.1 GB
Too big Get file →
IQ4_XS37.0 GB 38.8 GB
Too big Get file →
IQ4_NL39.1 GB 40.9 GB
Too big Get file →
Q4_K_S39.4 GB 41.2 GB
Too big Get file →
Q4_K_M41.5 GB 43.4 GB
Too big Get file →
Q5_K_S47.6 GB 49.4 GB
Too big Get file →
Q5_K_M48.9 GB 50.7 GB
Too big Get file →

For 8 GB VRAM, use IQ1_S (14.63 GB file, 17.8 GB memory at 8k context). For 16 GB VRAM, IQ1_M (16.04 GB file, 19.2 GB memory). For 24 GB VRAM, IQ2_S (21.46 GB file, 24.6 GB memory). Memory estimates are approximate.

What GPU do you need for tokyotech llm Swallow 70b instruct v0.1?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: no quant fits at an 8K context.
  • 16 GB VRAM: no quant fits at an 8K context.
  • 24 GB VRAM: the best quant that fits is IQ2_XS (20.4 GB file, about 23.6 GB in use).
  • 32 GB VRAM: the best quant that fits is IQ3_XS (28.4 GB file, about 31.6 GB in use).
  • 48 GB VRAM: the best quant that fits is Q4_K_M (41.5 GB file, about 44.7 GB in use).

Which tokyotech llm Swallow 70b instruct v0.1 quant should you choose?

Q4_K_M is the usual starting point for tokyotech llm Swallow 70b instruct v0.1: it balances file size and quality well. Here it is a 41.5 GB file.

If memory is tight, the smallest file is IQ1_S at 14.6 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturellama
Parameters69.2B
Layers80
Trained context4,096 tokens
KV cache (fp16)320 KB per token

Source

Quantized bymmnga
FamilyTokyotech
Licensellama2
Downloads (HF)2,815
Last updated2024-05-04

Frequently asked questions about tokyotech llm Swallow 70b instruct v0.1

How much VRAM does tokyotech llm Swallow 70b instruct v0.1 need?

About 44.7 GB at Q4_K_M with an 8K context window. That is the 41.5 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run tokyotech llm Swallow 70b instruct v0.1 on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 17.8 GB, so you would need a larger GPU or CPU offloading.

What is the file size of tokyotech llm Swallow 70b instruct v0.1 GGUF?

It depends on the quant: IQ1_S is 14.6 GB, IQ1_M is 16.0 GB, IQ2_XXS is 18.4 GB, IQ2_XS is 20.4 GB, IQ2_S is 21.5 GB, IQ2_M is 23.3 GB and 14 more.

What context length does tokyotech llm Swallow 70b instruct v0.1 support?

The GGUF header lists a trained context length of 4,096 tokens (4K). The KV cache grows with the context you use: about 2.7 GB at 8K and 1.3 GB at 4K.

Where can I download tokyotech llm Swallow 70b instruct v0.1 GGUF?

The files are on Hugging Face in the mmnga/tokyotech-llm-Swallow-70b-instruct-v0.1-gguf repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does tokyotech llm Swallow 70b instruct v0.1 use?

Hugging Face lists the license as llama2. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run this model on 8 GB VRAM?

Use the IQ1_S quantization, which requires 17.8 GB of VRAM at 8k context. It may still exceed 8 GB, so a higher VRAM GPU is recommended.

Does this model support vision?

No, the model has no vision capabilities.

What license does it use?

The model is licensed under llama2, which allows research use but restricts commercial deployment without permission.

Similar models

Browse all →

More from mmnga

All by mmnga →

More Tokyotech models

All Tokyotech →

Other large models

All large models →