diff --git a/global/default.nix b/global/default.nix
index 338fd46..1dfc2a6 100755
--- a/global/default.nix
+++ b/global/default.nix
@@ -33,13 +33,13 @@
         blasSupport = true;
       }).overrideAttrs
         (oldAttrs: rec {
-          version = "8184";
+          version = "8209";
           src = pkgs.fetchFromGitHub {
-            owner = "aagit";
+            owner = "ggml-org";
             repo = "llama.cpp";
-            # tag = "b${version}";
-            rev = "6ebf2e0d00d31acfc1a1fa9662e9a7d38bd07bf7"; # https://github.com/ggml-org/llama.cpp/pull/19970
-            hash = "sha256-xryajW0Cs1d+WDijspMTW21FDaZP9Grkb+uErMQCQ48=";
+            tag = "b${version}";
+            # rev = "6ebf2e0d00d31acfc1a1fa9662e9a7d38bd07bf7"; # https://github.com/ggml-org/llama.cpp/pull/19970
+            hash = "sha256-7z9mQZ/hgNS+doLCVPtax+FBhr6dEfmR9wZJTwtl/pM=";
             leaveDotGit = true;
             postFetch = ''
               git -C "$out" rev-parse --short HEAD > $out/COMMIT
@@ -54,7 +54,30 @@
           # for reproducible builds). We sacrifice portability for faster CPU layers.
           cmakeFlags = (oldAttrs.cmakeFlags or [ ]) ++ [
             "-DGGML_NATIVE=ON"
-            "-DCMAKE_CUDA_ARCHITECTURES=86" # RTX 3090 - needed since sandbox has no GPU
+            "-DGGML_LTO=ON" # Link Time Optimization for overall binary speed
+            "-DCMAKE_CUDA_ARCHITECTURES=86" # RTX 3090
+            "-DGGML_CUDA=ON"
+            "-DGGML_CUDA_FA=ON" # FlashAttention kernels (accelerated attention)
+            "-DGGML_CUDA_FA_ALL_QUANTS=ON" # Support for all KV cache quant types in FA
+            "-DGGML_CUDA_GRAPHS=ON" # CUDA Graphs for lower overhead inference
+            "-DGGML_CUDA_FORCE_CUBLAS=ON" # cuBLAS optimized prompt processing for Ampere+
+            "-DGGML_CUDA_PEER_MAX_BATCH_SIZE=256" # Increased for multi-GPU efficiency (split mode)
+            "-DGGML_CUDA_COMPRESSION_MODE=speed" # Fast binary loading (CUDA 12.8+)
+            "-DGGML_OPENMP=ON" # Optimal multi-threading on CPU
+            "-DGGML_LLAMAFILE=ON" # Use llamafile sgemm for faster CPU layers
+            "-DGGML_CPU_REPACK=ON" # Optimize Q4_0 quant handling
+            "-DGGML_AVX=ON"
+            "-DGGML_AVX2=ON"
+            "-DGGML_FMA=ON"
+            "-DGGML_F16C=ON"
+            "-DGGML_AVX512=ON" # Intel AVX-512 extensions
+            "-DGGML_AVX512_VNNI=ON" # Vector Neural Network Instructions
+            "-DGGML_AVX512_BF16=ON" # Bfloat16 support
+            "-DGGML_AVX_VNNI=ON" # VNNI for processors without AVX-512
+            "-DGGML_AMX_TILE=ON" # Intel Advanced Matrix Extensions (Sapphire Rapids+)
+            "-DGGML_AMX_INT8=ON"
+            "-DGGML_AMX_BF16=ON"
+            "-DGGML_BLAS=ON" # Uses internal BLAS provided by Nix (blasSupport=true works)
           ];
 
           # Disable Nix's NIX_ENFORCE_NO_NATIVE which strips -march=native flags
diff --git a/machines/homepc/configuration.nix b/machines/homepc/configuration.nix
index 8be6ba5..10cb166 100755
--- a/machines/homepc/configuration.nix
+++ b/machines/homepc/configuration.nix
@@ -87,7 +87,7 @@
   environment.etc."llama-swap/config.yaml".text = ''
     models:
       "Qwen3.5-35B-A3B-GGUF":
-        cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3.5-35B-A3B-GGUF:Q4_K_S --ctx-size 128000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -np 1  --cpu-moe --fit-target 256 --ubatch-size 1024 -fa on  --slots --slot-save-path /home/laythe/llamapcache  --jinja -kvu --no-mmproj --swa-checkpoints 32
+        cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3.5-35B-A3B-GGUF:UD-IQ4_XS --ctx-size 128000 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -np 1  --cpu-moe --fit-target 256 --ubatch-size 1024 -fa on  --slots --slot-save-path /home/laythe/llamapcache  --jinja -kvu --no-mmproj --swa-checkpoints 32
         ttl: 2400
       "Qwen3-1.7B-GGUF":
         cmd: llama-server --port ''${PORT} -hf unsloth/Qwen3-1.7B-GGUF:Q4_K_S --ctx-size 32768 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.00 -fa on --jinja -kvu