docs: add an experimental BF16 uncensored option (#1)

- docs: add an experimental BF16 uncensored option (73936dfd442fac196ab1108dbabbb9e43335e992)

README.md

@@ -25,6 +25,8 @@ tags:
 
 Measured on 17 and 18 October 2026 on one pair of DGX Spark. Every number on this page comes from a saved run in the [GitHub recipe](https://github.com/Elevatormusic/DeepSeek-V4.1-Flash-1017-EXL3-TensorFold-2x-DGX-Spark). The engine is [my TensorFold fork](https://github.com/Elevatormusic/TensorFold) (branch `deepseek-v41-tp2`).
 
+The speed results and the full 1M-token request test below apply to the base recipe. They do not apply to the optional uncensored BF16 overrides.
+
 ---
 
 ## At a glance
@@ -74,6 +76,18 @@ The Engram tables of the 1017 release are new. Do not use the tables of the olde
 
 ---
 
+## Optional uncensored BF16 overrides (experimental)
+
+This option adds 54 original dense BF16 matrices from the uncensored source to the base EXL3 model. It replaces `wo_b` and shared-expert `w2` in layers 10 through 36. It is not an all-BF16 checkpoint.
+
+The source is [dealignai/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED](https://huggingface.co/dealignai/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw-UNCENSORED/tree/89af9bf193575b7eeeb4f1dd85463ec24ea5bb4c), pinned to `89af9bf193575b7eeeb4f1dd85463ec24ea5bb4c`. See [recipe PR #14](https://github.com/Elevatormusic/DeepSeek-V4.1-Flash-1017-EXL3-TensorFold-2x-DGX-Spark/pull/14) and its [optional recipe instructions](https://github.com/Elevatormusic/DeepSeek-V4.1-Flash-1017-EXL3-TensorFold-2x-DGX-Spark/blob/00cf8f325fae83ab5805bb1d0a364ae696bde247/docs/UNCENSORED.md). This option needs the [engine change](https://github.com/Elevatormusic/TensorFold/pull/8) at `dccc4a997b7c18fc684781ee333273f8d902c510`. These pull requests are not merged. The default engine image does not support these overrides. The quick start below still uses the base model.
+
+Real replies passed with 8K-token and 64K-token prompts and with four concurrent requests. The server was configured for a 1M-token context, but no full 1M-token request was tested with these overrides. Both nodes logged `NV_ERR_NO_MEMORY` warnings while the API requests completed. No OOM kill was observed. This does not establish that memory failures are fixed.
+
+The 8 bpw override test produced a repetition loop. No decode-speed benefit was established, so BF16 is the documented option. These checks do not establish model quality or an uncensored benchmark score. See the [saved test limits](https://github.com/Elevatormusic/DeepSeek-V4.1-Flash-1017-EXL3-TensorFold-2x-DGX-Spark/blob/00cf8f325fae83ab5805bb1d0a364ae696bde247/runs/2026-10-19/uncensored-bf16/service-check.json). That summary records an earlier local engine, not GPU qualification of the new engine PR head or preparation wrapper.
+
+---
+
 ## What you need
 
 - **Two NVIDIA DGX Spark**, joined by one QSFP cable between their ConnectX-7 network ports.
@@ -205,5 +219,6 @@ The scripts, tests and documents in the recipe are Apache-2.0. The model, the ch
 | Recipe scripts, tests and documents | Apache-2.0 |
 | DeepSeek-V4.1-Flash-1017 (the model) | MIT |
 | LibertAIDAI EXL3 2.9 bpw weights | MIT |
+| Optional dealignai BF16 overrides | [DeepSeek V4.1 Flash 1017 licence](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/blob/main/LICENSE) |
 | TensorFold and the V4.1 engine | Apache-2.0 |
 | MiaAI-Lab vLLM kit and chat template | AGPL-3.0 |