c3: release notes -win variant

releases/llamafile-v0.10.3+opencoti.c3/RELEASE_NOTES.md

@@ -9,22 +9,39 @@ engine (#613 lane). Built from
 Version string: `opencoti-0.10.3-c3` — printed by `--version` and exposed
 at runtime.
 
-## Artifact
+## Artifacts
 
 | File | Size | sha256 |
 |---|---:|---|
 | `opencoti-llamafile-0.10.3-c3-x86_64.llamafile` | 4,812,296,839 B | `8a35b1f87b7bb7f2915d8756450f09008f4cd8331cf4a763e40edbcb8c4cf7e9` |
+| `opencoti-llamafile-0.10.3-c3-win-x86_64.llamafile.exe` | 55,429,522 B | `e5aa9a3d4e8274ca430f3a5557708955821cbc2d6d2c20f36a1c24228b0c642d` |
+| `dso/0.10.3-c3/ggml-cuda-x86_64.so` (side-load DSO, HF only) | 4,756,838,104 B | `5e0342ba9b2ded4847280067dfd6ebc4257e1b83a984881988c528c163b23e4a` |
 
-Embedded `ggml-cuda.so`: 4,756,838,104 B, sha256
-`5e0342ba9b2ded4847280067dfd6ebc4257e1b83a984881988c528c163b23e4a`
-(glibc floor GLIBC_2.29 — runs on Debian 11 / glibc 2.31 and newer).
+The main artifact's embedded `ggml-cuda.so` is byte-identical to the
+standalone side-load DSO above (glibc floor GLIBC_2.29 — runs on
+Debian 11 / glibc 2.31 and newer).
 
 Single-file Cosmopolitan APE. **Fat binary: runs natively on x86_64 and
 aarch64** (Linux/macOS/Windows/BSD) — on ARM64 hosts (DGX Spark, GH200,
-Apple Silicon) inference is CPU-only in this cut; the embedded CUDA DSO
-is x86_64-linux. GPU-enabled aarch64 (sbsa) is a planned separate
+Apple Silicon) inference is CPU-only in this cut; the CUDA DSO is
+x86_64-linux. GPU-enabled aarch64 (sbsa) is a planned separate
 artifact.
 
+### `-win` variant
+
+The full artifact cannot run on Windows: **Windows refuses to execute
+binaries larger than 4 GB**. `…-win-x86_64.llamafile.exe` is the same
+host APE (identical patches/features, `.exe` extension Windows
+requires) with **no embedded GPU backend** — 55 MB, CPU inference out
+of the box on Windows/Linux/macOS/BSD. GPU:
+
+- **Windows:** install CUDA Toolkit + MSVC; llamafile compiles
+  `ggml-cuda.dll` on the fly on first `-ngl` run (upstream mechanism).
+- **Linux x86_64:** side-load the standalone DSO:
+  `mkdir -p ~/.llamafile/v/0.10.3/ && cp ggml-cuda-x86_64.so
+  ~/.llamafile/v/0.10.3/ggml-cuda.so` (verify sha against
+  `SHA256SUMS` first).
+
 Run as a server:
 
 ```bash
@@ -38,6 +55,12 @@ Without `--server` the binary starts the interactive chat CLI.
 
 ## What's new vs c2 (2026-07-05)
 
+**Packaging:**
+- NEW `-win` variant (Windows-runnable bare APE, GPU via side-load)
+  + standalone side-load CUDA DSO published under `dso/` on HF
+- Supported/target model families (Gemma-4 primary, Qwen secondary)
+  now documented prominently in README + USAGE
+
 **Host binary (patches 0122–0133):**
 - RYS `--repeat-layers` weight-shared layer duplication + `--rys-probe`
   v2 task battery (#656–#673)