# Weight Reconstruction Fidelity on DeepSeek-V3 Weights

Source: [Weight Reconstruction Fidelity on DeepSeek-V3 Weights](<https://www.sota2.com/research/sota/weight-reconstruction-fidelity-on-deepseek-v3-weights>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-25T14:16:47.867Z

Evaluation of the fidelity of reconstructed model weights for the DeepSeek-V3 model, typically measured using L2 distance, sign agreement rate, and cosine similarity.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Weight Reconstruction Fidelity |
| Dataset (including any stated split) | DeepSeek-V3 Weights |
| Metric | Weight ΔW L2 Distance |
| Unit | SCALAR |
| Direction | Lower is better |
| Results loaded | 3 of 3 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [Weight ΔW L2 Distance](<https://www.sota2.com/research/sota/weight-reconstruction-fidelity-on-deepseek-v3-weights.md?metric=Weight%20%CE%94W%20L2%20Distance>): Lower is better; unit: SCALAR.

- [Sign Agreement Rate (%)](<https://www.sota2.com/research/sota/weight-reconstruction-fidelity-on-deepseek-v3-weights.md?metric=Sign%20Agreement%20Rate%20(%25)>): Higher is better; unit: PERCENT.

- [Cosine Similarity](<https://www.sota2.com/research/sota/weight-reconstruction-fidelity-on-deepseek-v3-weights.md?metric=Cosine%20Similarity>): Higher is better; unit: SCALAR.

## Recorded results

| Method | Weight ΔW L2 Distance | Result date | Report paper |
| --- | --- | --- | --- |
| Post-trained | 0 | Not recorded | 69c36c190636e5bf4afc651f |
| AbsMax | 48641 | Not recorded | 69c36c190636e5bf4afc651f |
| AbsMax | 51790 | Not recorded | 69c36c190636e5bf4afc651f |

### Post-trained

Weight ΔW L2 Distance: 0. Result ID: 69c3ece347f34ce924adad6f.

[Reporting paper](<https://www.sota2.com/research/paper/69c36c190636e5bf4afc651f.md>)

Recorded evaluation configuration: {"Precision":"BF16"}. This record does not assert independent reproduction.

### AbsMax

Weight ΔW L2 Distance: 48641. Result ID: 69c3ece347f34ce924adad70.

[Reporting paper](<https://www.sota2.com/research/paper/69c36c190636e5bf4afc651f.md>)

Recorded evaluation configuration: {"Precision":"FP8","Quantization Granularity":"block"}. This record does not assert independent reproduction.

### AbsMax

Weight ΔW L2 Distance: 51790. Result ID: 69c3ece347f34ce924adad71.

[Reporting paper](<https://www.sota2.com/research/paper/69c36c190636e5bf4afc651f.md>)

Recorded evaluation configuration: {"Precision":"FP8","Quantization Granularity":"channel"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.