license: apache-2.0
base_model: MagicalAlchemist/Qwen3-1.7B-Magic_decensored
base_model_relation: quantized
library_name: gguf
pipeline_tag: text-generation
tags:
- gguf
- tpn-workbench
- llama.cpp
- gratisai
Qwen3-1.7B-Magic_decensored · TPN-optimised GGUF
Quantised and measured by TPN Workbench from MagicalAlchemist/Qwen3-1.7B-Magic_decensored @ b5727bf. Every format is scored against the BF16 reference on next-token KL divergence, perplexity and top-1 agreement (llama.cpp b11528).
Full report: https://bench.trueperformancenetwork.com/r/hCN5YMVh_9qQWSgezAelvQ
TPN recommendation: Q4_K_M: Highest-fidelity format that fits 1.2 GB: 1.11 GB, closeness 0.059.
Files
| file | format | size | mean KLD | KL similarity | PPL Δ | top-1 | |
|---|---|---|---|---|---|---|---|
Qwen3-1.7B-Magic_decensored-Q4_K_M.gguf |
Q4_K_M | 1.11 GB | 0.0586 | 94.3 % | +6.22 % | 90.5 % | TPN recommendation |
Qwen3-1.7B-Magic_decensored-Q5_0-tuned.gguf |
Q5_0+ | 1.23 GB | 0.0283 | 97.2 % | +2.25 % | 93.6 % | |
Qwen3-1.7B-Magic_decensored-Q8_0.gguf |
Q8_0 | 1.83 GB | 0.0034 | 99.7 % | +1.41 % | 97.5 % |
Q5_0+ is a TPN fidelity-tuned format: starting from Q5_0, TPN picks each tensor's quantisation type by measurement, keeping the file small while staying closer to the original than the plain format.
Lower KLD = closer to the original. K-quants (Q4_K_M, Q8_0) have SIMD kernels in llama.cpp WebAssembly, so they are the faster pick in the browser; IQ* formats are smallest-for-quality on native CPU/GPU builds.
Measurements
| format | size | bpw | mean KLD | PPL Δ | top-1 |
|---|---|---|---|---|---|
| Q8_0 | 1.83 GB | 8.5 | 0.0034 | +1.41 % | 97.5 % |
| Q5_0 | 1.31 GB | 5.5 | 0.0544 | +4.80 % | 91.0 % |
| Q5_0+ | 1.23 GB | 5.5 | 0.0283 | +2.25 % | 93.6 % |
| Q4_0M+ | 1.20 GB | 4.92 | 0.0738 | -6.23 % | 89.5 % |
| Q4_0 | 1.13 GB | 4.5 | 0.1687 | -0.08 % | 84.1 % |
| Q4_K_M | 1.11 GB | 4.8 | 0.0586 | +6.22 % | 90.5 % |
| IQ4_XS | 1.01 GB | 4.25 | 0.0761 | +2.16 % | 88.9 % |
| Q3_K_M | 0.94 GB | 3.85 | 0.1705 | +5.12 % | 84.0 % |
| Q2_K | 0.78 GB | 2.93 | 0.4874 | +18.21 % | 73.1 % |
Task benchmarks
Score in % (TPN ran no BF16 baseline for comparison).
| format | gsm8k | ifeval |
|---|---|---|
| Q4_K_M | 45.6 | 38.3 |
Use
llama-server -hf tpnlabs/Qwen3-1.7B-Magic_decensored-TPN-GGUF --hf-file Qwen3-1.7B-Magic_decensored-Q4_K_M.gguf
Runs fully on-device in Gratis, in the browser or desktop app.
Licence: inherited from the base model (apache-2.0).