SM

SmallThinker 21BA3B Instruct

SmallthinkerGGUF by gabriellarsonUpdated 2025-07-28
Download on Hugging Face
Parameters21.5B
Context16K
Layers52
Architecturesmallthinker
Smallest quant7.5 GB
Quick answer

SmallThinker 21BA3B Instruct needs about 14.5 GB of memory at Q4_K_M (a 13.1 GB file) with an 8K context window, so it fits on a 16 GB GPU. This repository has 18 quants, from 7.5 GB (Q2_K_S) to 43.0 GB (F16).

21.5B parameter Mixture-of-Experts LLM for local deployment

SmallThinker 21BA3B Instruct is a 21.5B parameter Mixture-of-Experts model designed for on‑device inference. It supports 16384 token context and is trained mainly on English. The model is released under the apache-2.0 license.

Good at

  • On‑device inference
  • Mixture‑of‑Experts architecture
  • Large context length

Limits

  • Requires 8k context for memory estimates
  • No vision support
  • License is apache-2.0

Best for

  • Local chatbot applications
  • Edge device AI
  • Privacy‑focused deployments

Who it’s for. Developers building AI applications that must run entirely on local hardware without cloud access. The model is suitable for devices with limited GPU memory and slow storage, thanks to its sparse MoE design.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q2_K_S7.5 GB 8.9 GB
Too big Get file →
Q2_K8.0 GB 9.4 GB
Too big Get file →
IQ3_XS9.0 GB 10.3 GB
Too big Get file →
Q3_K_S9.4 GB 10.8 GB
Too big Get file →
IQ3_S9.4 GB 10.8 GB
Too big Get file →
IQ3_M9.6 GB 11.0 GB
Too big Get file →
Q3_K_M10.4 GB 11.8 GB
Too big Get file →
Q3_K_L11.3 GB 12.6 GB
Too big Get file →
IQ4_XS11.6 GB 13.0 GB
Too big Get file →
Q4_012.3 GB 13.7 GB
Too big Get file →
Q4_K_S12.3 GB 13.7 GB
Too big Get file →
Q4_K_M13.1 GB 14.5 GB
Too big Get file →
Q5_K_S14.9 GB 16.2 GB
Too big Get file →
Q5_014.9 GB 16.3 GB
Too big Get file →
Q5_K_M15.3 GB 16.7 GB
Too big Get file →
Q6_K17.7 GB 19.0 GB
Too big Get file →
Q8_022.9 GB 24.3 GB
Too big Get file →
F1643.0 GB 44.4 GB
Too big Get file →

For 8 GB VRAM, use Q4_K_M (13.09 GB file, 14.5 GB memory at 8k context). For 16 GB VRAM, Q5_K_M (15.31 GB file, 16.7 GB memory) is a good balance. If more memory is available, Q6_K (17.68 GB file, 19 GB memory) offers higher precision.

What GPU do you need for SmallThinker 21BA3B Instruct?

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: the best quant that fits is Q3_K_M (10.4 GB file, about 11.8 GB in use).
  • 16 GB VRAM: the best quant that fits is Q4_K_M (13.1 GB file, about 14.5 GB in use).
  • 24 GB VRAM: the best quant that fits is Q6_K (17.7 GB file, about 19.0 GB in use).
  • 32 GB VRAM: the best quant that fits is Q8_0 (22.9 GB file, about 24.3 GB in use).
  • 48 GB VRAM: the best quant that fits is F16 (43.0 GB file, about 44.4 GB in use).

Which SmallThinker 21BA3B Instruct quant should you choose?

Q4_K_M is the usual starting point for SmallThinker 21BA3B Instruct: it balances file size and quality well. Here it is a 13.1 GB file.

The Q8_0 file is 22.9 GB, about 1.7 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is Q2_K_S at 7.5 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

Architecturesmallthinker
Parameters21.5B (approximate, mixture of experts)
Layers52
Trained context16,384 tokens
KV cache (fp16)104 KB per token

Source

Quantized bygabriellarson
Licenseapache-2.0
Downloads (HF)409
Last updated2025-07-28

Frequently asked questions about SmallThinker 21BA3B Instruct

How much VRAM does SmallThinker 21BA3B Instruct need?

About 14.5 GB at Q4_K_M with an 8K context window. That is the 13.1 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 SmallThinker 21BA3B Instruct on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 8.9 GB, so you would need a larger GPU or CPU offloading.

What is the file size of SmallThinker 21BA3B Instruct GGUF?

It depends on the quant: Q2_K_S is 7.5 GB, Q2_K is 8.0 GB, IQ3_XS is 9.0 GB, Q3_K_S is 9.4 GB, IQ3_S is 9.4 GB, IQ3_M is 9.6 GB and 12 more.

What context length does SmallThinker 21BA3B Instruct support?

The GGUF header lists a trained context length of 16,384 tokens (16K). The KV cache grows with the context you use: about 0.9 GB at 8K and 1.7 GB at 16K.

Where can I download SmallThinker 21BA3B Instruct GGUF?

The files are on Hugging Face in the gabriellarson/SmallThinker-21BA3B-Instruct-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 SmallThinker 21BA3B Instruct use?

Hugging Face lists the license as apache-2.0. 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 SmallThinker 21BA3B Instruct on 8 GB VRAM?

Yes, the Q4_K_M quantization requires about 14.5 GB memory at 8k context, which fits within 8 GB VRAM when using efficient memory management.

Does the model support vision tasks?

No, SmallThinker 21BA3B Instruct does not include vision capabilities.

What is the context length of the model?

The model supports a maximum context length of 16384 tokens.

Other GGUFs of SmallThinker-21BA3B-Instruct

Similar models

Browse all →

More from gabriellarson

All by gabriellarson →

More Smallthinker models

All Smallthinker →

Other medium models

All medium models →