Quantization publisher
mmnga
mmnga publishes 69 quantizations across 3 models in our index, averaging 4.172 effective bits per weight. Their files differ in size from other publishers' builds of the same nominal quantization on 25 of the pairs we can compare — the same label does not mean the same file.
From the file· summed file bytes
Repositories
3
Quantizations
69
Models covered
3
Avg effective bpw
4.172
across their files
Same model, same quant label, different bytes
largest disagreements first
| Model | Quant | mmnga | vs | Theirs | Difference |
|---|---|---|---|---|---|
| Llama-3_3-Nemotron-Super-49B-v1 | Q4_0 | 26.39 GiB | bartowski | 26.50 GiB | -0.4% |
| Llama-3_3-Nemotron-Super-49B-v1 | IQ4_XS | 25.03 GiB | unsloth | 25.06 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | Q5_0 | 10.28 GiB | gabriellarson | 10.30 GiB | -0.2% |
| Moonlight-16B-A3B-Instruct | Q4_0 | 8.46 GiB | gabriellarson | 8.48 GiB | -0.3% |
| Moonlight-16B-A3B-Instruct | IQ2_XXS | 5.37 GiB | gabriellarson | 5.38 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | IQ2_XS | 5.67 GiB | gabriellarson | 5.68 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | IQ2_S | 5.72 GiB | gabriellarson | 5.73 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | Q6_K | 13.29 GiB | gabriellarson | 13.30 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | IQ2_M | 6.03 GiB | gabriellarson | 6.03 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | Q2_K | 6.12 GiB | gabriellarson | 6.13 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | IQ3_XXS | 6.62 GiB | gabriellarson | 6.62 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | Q3_K_S | 7.12 GiB | gabriellarson | 7.12 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q3_K_M | 7.72 GiB | gabriellarson | 7.72 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q3_K_L | 8.03 GiB | gabriellarson | 8.03 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | IQ3_XS | 6.78 GiB | gabriellarson | 6.78 GiB | -0.1% |
| Moonlight-16B-A3B-Instruct | IQ3_S | 7.12 GiB | gabriellarson | 7.12 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | IQ3_M | 7.18 GiB | gabriellarson | 7.19 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q4_K_M | 9.81 GiB | gabriellarson | 9.82 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q4_K_S | 9.04 GiB | gabriellarson | 9.05 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q5_K_S | 10.56 GiB | gabriellarson | 10.56 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q5_K_M | 11.21 GiB | gabriellarson | 11.21 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | IQ4_XS | 8.14 GiB | gabriellarson | 8.15 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | Q8_0 | 15.81 GiB | gabriellarson | 15.81 GiB | -0.0% |
| Moonlight-16B-A3B-Instruct | IQ4_NL | 8.46 GiB | gabriellarson | 8.46 GiB | -0.0% |
| Llama-3_3-Nemotron-Super-49B-v1 | Q2_K | 17.45 GiB | unsloth | 17.45 GiB | -0.0% |
A quantization label describes a target, not a recipe. Publishers make different choices about which tensors to keep at higher precision, and some apply an importance matrix while others don't — so two files both honestly labelled the same thing can differ measurably in size and in quality.
Models they publish
| Model | Quantizations | Smallest |
|---|---|---|
| Moonlight-16B-A3B-Instruct | 24 | 4.77 GiB |
| Llama-3.1-Swallow-8B-Instruct-v0.5 | 22 | 2.23 GiB |
| Llama-3_3-Nemotron-Super-49B-v1 | 23 | 10.27 GiB |