Add three APEX GGUF options to the checkpoint explorer

README.md

@@ -10,6 +10,7 @@ license: apache-2.0
 models:
 - Accio-Lab/occamy-1.0
 - Accio-Lab/occamy-1.0-GGUF
+- Accio-Lab/occamy-1.0-APEX-GGUF
 - Accio-Lab/occamy-1.0-FP8
 - Accio-Lab/occamy-1.0-NVFP4
 - Accio-Lab/occamy-1.0-MLX-8bit
@@ -22,7 +23,7 @@ short_description: Compare Occamy weights and copy deployment commands.
 
 # Occamy Checkpoint Explorer
 
-The official Occamy model selection, download and deployment guide. Compare seventeen GGUF quantizations, BF16/FP8/NVFP4 server checkpoints, MLX 3/4/6/8-bit candidates and the experimental MTP head.
+The official Occamy model selection, download and deployment guide. Compare twenty GGUF options (seventeen standard recipes and three APEX mixed-precision profiles), BF16/FP8/NVFP4 server checkpoints, MLX 3/4/6/8-bit candidates and the experimental MTP head.
 
 This static Space runs in the browser with no inference backend or paid compute. It provides links and deployment recipes; it does not run the 35B model or download weights automatically.
 

app.js

@@ -8,7 +8,7 @@ let mlxRecipe = 'generate';
 let currentApiRequest = '';
 let toastTimer;
 const notes = {
-  gguf: 'Seventeen GGUF quantizations for llama.cpp. All files use qwen2 metadata; normalize prompt content to NFC. Vision requires the separate projector. Recent additions have native qwen2 sample checks; earlier benchmarks used the documented qwen35 protocol.',
+  gguf: 'Twenty GGUF options: seventeen standard recipes plus three APEX MoE mixed-precision profiles. All files use qwen2 metadata; normalize input to NFC. Vision uses the separate projector. Read the validation scope for each option.',
   nvidia: 'Full-model checkpoints for GPU servers. File sizes include preserved vision weights. NVFP4 on H200 uses Marlin W4A16; native FP4 requires compatible hardware and kernels. Adjust context and memory settings to your GPU.',
   mlx: 'Text-only native MLX candidates. Linux native validation passed; Mac Metal acceptance is pending. The recipes below await Mac validation. Allow extra unified memory beyond the weight files.',
   research: 'Experimental MTP draft head, not a standalone model. Download a base separately and follow the patched SGLang setup. Performance after the final MTP3 fix has not been measured.'

catalog.json

@@ -80,6 +80,22 @@
       "url": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF",
       "download": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF/resolve/main/occamy-1.0-IQ3_M.gguf?download=true"
     },
+    {
+      "id": "apex-compact",
+      "label": "APEX Compact",
+      "repo": "Accio-Lab/occamy-1.0-APEX-GGUF",
+      "revision": "d5c04369b99f0de7007d2cc351b0206b72912f32",
+      "platform": "gguf",
+      "description": "Upstream APEX Compact MoE mixed-precision allocation, generated directly from BF16 without an importance matrix. Normalize prompt content to NFC.",
+      "status": "Published",
+      "command": "hf download Accio-Lab/occamy-1.0-APEX-GGUF occamy-1.0-APEX-Compact.gguf --local-dir ./occamy-gguf\n\nllama-server -m ./occamy-gguf/occamy-1.0-APEX-Compact.gguf \\\n  -ngl 999 -c 8192 -np 1 -fa on --jinja \\\n  --host 127.0.0.1 --port 8000 --alias occamy",
+      "evidence": "B200 native qwen2 + NFC: 20/20 authored functional cases, including executed Python and tool round trips; subset PPL 8.4567. All 733 tensors finite and 16/16 tokenizer probes. Bounded checks; image, CPU and Mac execution were not tested.",
+      "file": "occamy-1.0-APEX-Compact.gguf",
+      "bytes": 16538851072,
+      "vision": true,
+      "url": "https://huggingface.co/Accio-Lab/occamy-1.0-APEX-GGUF",
+      "download": "https://huggingface.co/Accio-Lab/occamy-1.0-APEX-GGUF/resolve/main/occamy-1.0-APEX-Compact.gguf?download=true"
+    },
     {
       "id": "q3_k_m",
       "label": "Q3_K_M",
@@ -192,6 +208,22 @@
       "url": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF",
       "download": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF/resolve/main/occamy-1.0-Q4_K_M.gguf?download=true"
     },
+    {
+      "id": "apex-quality",
+      "label": "APEX Quality",
+      "repo": "Accio-Lab/occamy-1.0-APEX-GGUF",
+      "revision": "d5c04369b99f0de7007d2cc351b0206b72912f32",
+      "platform": "gguf",
+      "description": "Upstream APEX Quality MoE mixed-precision allocation, generated directly from BF16 without an importance matrix. Normalize prompt content to NFC.",
+      "status": "Published",
+      "command": "hf download Accio-Lab/occamy-1.0-APEX-GGUF occamy-1.0-APEX-Quality.gguf --local-dir ./occamy-gguf\n\nllama-server -m ./occamy-gguf/occamy-1.0-APEX-Quality.gguf \\\n  -ngl 999 -c 8192 -np 1 -fa on --jinja \\\n  --host 127.0.0.1 --port 8000 --alias occamy",
+      "evidence": "B200 native qwen2 + NFC: 19/20 authored functional cases, including executed Python and tool round trips; subset PPL 8.2568. All 733 tensors finite and 16/16 tokenizer probes. Bounded checks; image, CPU and Mac execution were not tested.",
+      "file": "occamy-1.0-APEX-Quality.gguf",
+      "bytes": 22819401472,
+      "vision": true,
+      "url": "https://huggingface.co/Accio-Lab/occamy-1.0-APEX-GGUF",
+      "download": "https://huggingface.co/Accio-Lab/occamy-1.0-APEX-GGUF/resolve/main/occamy-1.0-APEX-Quality.gguf?download=true"
+    },
     {
       "id": "q5_k_s",
       "label": "Q5_K_S",
@@ -240,6 +272,22 @@
       "url": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF",
       "download": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF/resolve/main/occamy-1.0-Q5_K_M.gguf?download=true"
     },
+    {
+      "id": "apex-balanced",
+      "label": "APEX Balanced",
+      "repo": "Accio-Lab/occamy-1.0-APEX-GGUF",
+      "revision": "d5c04369b99f0de7007d2cc351b0206b72912f32",
+      "platform": "gguf",
+      "description": "Upstream APEX Balanced MoE mixed-precision allocation, generated directly from BF16 without an importance matrix. Normalize prompt content to NFC.",
+      "status": "Published",
+      "command": "hf download Accio-Lab/occamy-1.0-APEX-GGUF occamy-1.0-APEX-Balanced.gguf --local-dir ./occamy-gguf\n\nllama-server -m ./occamy-gguf/occamy-1.0-APEX-Balanced.gguf \\\n  -ngl 999 -c 8192 -np 1 -fa on --jinja \\\n  --host 127.0.0.1 --port 8000 --alias occamy",
+      "evidence": "B200 native qwen2 + NFC: 19/20 authored functional cases, including executed Python and tool round trips; subset PPL 8.2994. All 733 tensors finite and 16/16 tokenizer probes. Bounded checks; image, CPU and Mac execution were not tested.",
+      "file": "occamy-1.0-APEX-Balanced.gguf",
+      "bytes": 25335983872,
+      "vision": true,
+      "url": "https://huggingface.co/Accio-Lab/occamy-1.0-APEX-GGUF",
+      "download": "https://huggingface.co/Accio-Lab/occamy-1.0-APEX-GGUF/resolve/main/occamy-1.0-APEX-Balanced.gguf?download=true"
+    },
     {
       "id": "q6_k",
       "label": "Q6_K",
@@ -411,4 +459,4 @@
     "bytes": 899282944,
     "url": "https://huggingface.co/Accio-Lab/occamy-1.0-GGUF/resolve/main/mmproj-occamy-1.0-F16.gguf?download=true"
   }
-}
+}

index.html

@@ -20,7 +20,7 @@
         <h1 aria-label="Occamy-1.0"><span class="hero-occamy-wordmark"><img src="https://accio-lab.github.io/occamy/brand/occamy.png" alt="" aria-hidden="true"><span>ccamy-1.0</span></span></h1>
         <p class="hero-subtitle">Choose a checkpoint.<br>Run Occamy on your hardware.</p>
         <p class="hero-deck">Compare weight files, download an official release, and copy the command for your runtime.</p>
-        <div class="hero-facts"><span><b>35B</b> total parameters</span><span><b>3B</b> active per token</span><span><b>17</b> GGUF options</span><span><b>4</b> MLX candidates</span></div>
+        <div class="hero-facts"><span><b>35B</b> total parameters</span><span><b>3B</b> active per token</span><span><b>20</b> GGUF options</span><span><b>4</b> MLX candidates</span></div>
         <div class="hero-actions"><a class="button button-primary" href="#explorer-title">Choose a checkpoint <span aria-hidden="true">↓</span></a><a class="button button-secondary" href="https://huggingface.co/papers/2609.11977" target="_blank" rel="noopener noreferrer">Read the paper ↗</a></div>
       </div>
     </section>