NU

NuminaMath-7B-TIR

NuminamathGGUF by bartowskiUpdated 2024-07-10
Download on Hugging Face
Parameters6.91B
Context4K
Layers30
Architecturellama
Smallest quant2.2 GB
Quick answer

NuminaMath-7B-TIR needs about 8.7 GB of memory at Q4_K_M (a 4.2 GB file) with an 8K context window, so it fits on a 12 GB GPU. This repository has 21 quants, from 2.2 GB (IQ2_XS) to 27.6 GB (F32).

NuminaMath-7B-TIR: a 7B parameter math LLM fine‑tuned for tool‑integrated reasoning

NuminaMath-7B-TIR is a 6.91B parameter language model trained to solve mathematical problems using chain‑of‑thought and Python REPL. It is a fine‑tuned version of deepseek‑math‑7b‑base and is released under Apache 2.0.

Good at

  • Math problem solving
  • Tool‑integrated reasoning
  • Chain‑of‑thought generation

Limits

  • No vision support
  • Requires 8k context for full performance
  • License is Apache 2.0

Best for

  • Solving math Olympiad problems
  • Generating Python code for intermediate calculations
  • Educational math tutoring

Who it’s for. Researchers and developers needing a math‑focused LLM that can execute Python code for intermediate results. The model is suitable for academic projects and educational tools where mathematical reasoning is required.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ2_XS2.2 GB 4.7 GB
Fits Get file →
IQ2_S2.4 GB 4.9 GB
Fits Get file →
IQ2_M2.5 GB 5.1 GB
Fits Get file →
Q2_K2.7 GB 5.2 GB
Fits Get file →
IQ3_XXS2.8 GB 5.3 GB
Fits Get file →
IQ3_XS3.0 GB 5.5 GB
Fits Get file →
Q3_K_S3.1 GB 5.7 GB
Fits Get file →
IQ3_M3.3 GB 5.8 GB
Fits Get file →
Q3_K_M3.5 GB 6.0 GB
Fits Get file →
Q3_K_L3.7 GB 6.3 GB
Fits Get file →
IQ4_XS3.8 GB 6.3 GB
Fits Get file →
Q4_K_S4.0 GB 6.5 GB
Fits Get file →
Q4_K_M4.2 GB 6.7 GB
Fits Get file →
Q4_K_L4.5 GB 7.0 GB
Fits Get file →
Q5_K_S4.8 GB 7.3 GB
Tight Get file →
Q5_K_M4.9 GB 7.4 GB
Tight Get file →
Q5_K_L5.2 GB 7.7 GB
Tight Get file →
Q6_K5.7 GB 8.2 GB
Too big Get file →
Q6_K_L5.9 GB 8.4 GB
Too big Get file →
Q8_07.3 GB 9.9 GB
Too big Get file →
F3227.6 GB 30.2 GB
Too big Get file →

For 8 GB VRAM, use Q4_K_S (file size 4.03 GB, memory 8.6 GB) or Q4_K_M (4.22 GB, 8.7 GB). For 12 GB VRAM, Q5_K_S (4.81 GB, 9.3 GB) or Q5_K_M (4.93 GB, 9.5 GB) are suitable. For 24 GB VRAM, Q6_K (5.67 GB, 10.2 GB) or Q6_K_L (5.88 GB, 10.4 GB) provide higher quality.

What GPU do you need for NuminaMath-7B-TIR?

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: the best quant that fits is Q3_K_M (3.5 GB file, about 8.0 GB in use).
  • 12 GB VRAM: the best quant that fits is Q8_0 (7.3 GB file, about 11.9 GB in use).
  • 16 GB VRAM: the best quant that fits is Q8_0 (7.3 GB file, about 11.9 GB in use).
  • 24 GB VRAM: the best quant that fits is Q8_0 (7.3 GB file, about 11.9 GB in use).
  • 32 GB VRAM: the best quant that fits is Q8_0 (7.3 GB file, about 11.9 GB in use).
  • 48 GB VRAM: the best quant that fits is F32 (27.6 GB file, about 32.2 GB in use).

Which NuminaMath-7B-TIR quant should you choose?

Q4_K_M is the usual starting point for NuminaMath-7B-TIR: it balances file size and quality well. Here it is a 4.2 GB file.

The Q8_0 file is 7.3 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 IQ2_XS at 2.2 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
Parameters6.91B
Layers30
Trained context4,096 tokens
KV cache (fp16)480 KB per token

Source

Quantized bybartowski
Licenseapache-2.0
Downloads (HF)839
Last updated2024-07-10

Frequently asked questions about NuminaMath-7B-TIR

How much VRAM does NuminaMath-7B-TIR need?

About 8.7 GB at Q4_K_M with an 8K context window. That is the 4.2 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 NuminaMath-7B-TIR on an 8 GB GPU?

Yes, with Q3_K_M: the 3.5 GB file needs about 8.0 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of NuminaMath-7B-TIR GGUF?

It depends on the quant: IQ2_XS is 2.2 GB, IQ2_S is 2.4 GB, IQ2_M is 2.5 GB, Q2_K is 2.7 GB, IQ3_XXS is 2.8 GB, IQ3_XS is 3.0 GB and 15 more.

What context length does NuminaMath-7B-TIR support?

The GGUF header lists a trained context length of 4,096 tokens (4K). The KV cache grows with the context you use: about 4.0 GB at 8K and 2.0 GB at 4K.

Where can I download NuminaMath-7B-TIR GGUF?

The files are on Hugging Face in the bartowski/NuminaMath-7B-TIR-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 NuminaMath-7B-TIR 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 NuminaMath-7B-TIR on 8 GB VRAM?

Yes, use a K‑quant such as Q4_K_S or Q4_K_M, which fit within 8 GB VRAM and provide acceptable performance.

What context length does the model support?

The model was trained on 4096 tokens and performs best with an 8k context length.

Is the model licensed for commercial use?

The model is released under Apache 2.0, which allows commercial use with attribution.

Other GGUFs of NuminaMath-7B-TIR

Similar models

Browse all →

More from bartowski

All by bartowski →

More Numinamath models

All Numinamath →

Other small models

All small models →