Add official FP8 release to checkpoint list

README.md

@@ -3,6 +3,8 @@ license: apache-2.0
 library_name: transformers
 pipeline_tag: image-text-to-text
 base_model: Qwen/Qwen3.6-35B-A3B
+datasets:
+- Accio-Lab/occamy-data-1.0
 tags:
 - agent
 - agentic
@@ -15,11 +17,11 @@ tags:
 
 <div align="center">
   <picture>
-    <img src="assets/accio.png" width="34%" alt="Accio">
+    <img src="https://raw.githubusercontent.com/Accio-Lab/occamy/9818009b54abbbc5f3c5620c93dfbf7e5b90dd06/assets/accio.svg" width="34%" alt="Accio">
   </picture>
   &nbsp;&nbsp;&nbsp;&nbsp;
   <picture>
-    <img src="assets/occamy.png" width="13%" alt="Occamy logo">
+    <img src="https://raw.githubusercontent.com/Accio-Lab/occamy/9818009b54abbbc5f3c5620c93dfbf7e5b90dd06/assets/occamy.png" width="13%" alt="Occamy logo">
   </picture>
   <h1>Occamy-1.0</h1>
   <p><strong>Open Pareto-frontier 35B Intelligence for Co-work</strong></p>
@@ -30,16 +32,21 @@ tags:
 <div align="center" style="line-height: 1;">
   <a href="https://accio-lab.github.io/occamy/"><img alt="Project Website" src="https://img.shields.io/badge/Website-Occamy--1.0-087F6A"></a>
   <a href="https://huggingface.co/Accio-Lab/Occamy-1.0"><img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Model-Occamy--1.0-FFD21E"></a>
-  <a href="https://github.com/Accio-Lab/Dressage"><img alt="Dressage" src="https://img.shields.io/badge/Training-Dressage-087F6A"></a>
   <a href="https://arxiv.org/pdf/2609.11977"><img alt="Technical Report" src="https://img.shields.io/badge/%F0%9F%93%84%20Report-Occamy--1.0-B31B1B"></a>
+  <a href="https://github.com/Accio-Lab/Dressage"><img alt="Dressage" src="https://img.shields.io/badge/Training-Dressage-087F6A"></a>
   <a href="LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Apache%202.0-blue"></a>
 </div>
 
-<p align="center">
-  <a href="https://accio-lab.github.io/occamy/">Project Website</a> &nbsp;|&nbsp;
-  <a href="https://huggingface.co/Accio-Lab/Occamy-1.0">Model Weights</a> &nbsp;|&nbsp;
-  <a href="https://github.com/Accio-Lab/Dressage">Training Framework</a>
-</p>
+## Available Checkpoints
+
+| Format | Repository | Notes |
+|---|---|---|
+| Original BF16 | [Accio-Lab/occamy-1.0](https://huggingface.co/Accio-Lab/occamy-1.0) | Original model weights and configuration. |
+| GGUF | [Accio-Lab/occamy-1.0-GGUF](https://huggingface.co/Accio-Lab/occamy-1.0-GGUF) | Q4_K_M, Q8_0, and the matching F16 vision projector. Normalize text inputs to NFC; see the model card for tokenizer compatibility. |
+| FP8 | [Accio-Lab/occamy-1.0-FP8](https://huggingface.co/Accio-Lab/occamy-1.0-FP8) | Blockwise FP8 weights with dynamic activation quantization; see the model card for serving instructions and paired BF16 evaluation. |
+| NVFP4 | [Accio-Lab/occamy-1.0-NVFP4](https://huggingface.co/Accio-Lab/occamy-1.0-NVFP4) | Mixed-precision ModelOpt checkpoint with NVFP4 routed experts. See its model card for the tested runtime and hardware requirements. |
+
+The quantized releases include bounded regression checks, not a re-run of the full benchmarks below. They do not include MTP weights; MTP combinations are being evaluated separately.
 
 ## 1. Model Introduction
 
@@ -82,75 +89,51 @@ Architecture fields follow the starting checkpoint's published model card. Occam
 
 <div align="center">
   <picture>
-    <img src="assets/occamy-main-results.png" width="100%" alt="Occamy-1.0 results on co-work, tool-use, coding, and business benchmarks">
-  </picture>
-</div>
-
-### Full Evaluation
-
-<table>
-  <thead>
-    <tr>
-      <th rowspan="2" align="left" bgcolor="#EAF6F3">Benchmark</th>
-      <th colspan="6" align="center" bgcolor="#EAF6F3"><font color="#087F6A">35B-A3B Models</font></th>
-      <th colspan="4" align="center" bgcolor="#EAF6F3"><font color="#087F6A">Large-scale Models</font></th>
-    </tr>
-    <tr>
-      <th align="center" bgcolor="#087F6A"><font color="#FFFFFF">Occamy-1.0</font></th>
-      <th align="center">Qwen3.6<br>35B-A3B</th>
-      <th align="center">Agents-A1</th>
-      <th align="center">Nex-N2-mini</th>
-      <th align="center">BigBang-1.0</th>
-      <th align="center">Ornith-1.5</th>
-      <th align="center">GPT-5.6 Sol</th>
-      <th align="center">Qwen3.8-Max</th>
-      <th align="center">DeepSeek V4<br>Pro (0813)</th>
-      <th align="center">GLM-5.2</th>
-    </tr>
-  </thead>
-  <tbody>
-    <tr><td colspan="11" bgcolor="#DDF1EC"><font color="#087F6A"><strong>Co-work</strong></font></td></tr>
-    <tr><td>Claw-Eval (average)</td><td align="right"><strong>82.20</strong></td><td align="right">69.50</td><td align="right"><u>69.90</u></td><td align="right">66.60</td><td align="right">63.50</td><td align="right">64.40</td><td align="right"><u>81.80</u></td><td align="right"><strong>83.92</strong></td><td align="right">81.70</td><td align="right">81.60</td></tr>
-    <tr><td>Claw-Eval (Pass³)</td><td align="right"><strong>71.40</strong></td><td align="right"><u>54.80</u></td><td align="right">41.70</td><td align="right">37.00</td><td align="right">40.20</td><td align="right">48.70</td><td align="right">68.90</td><td align="right"><u>73.68</u></td><td align="right"><strong>74.50</strong></td><td align="right">68.30</td></tr>
-    <tr><td>WildClawBench</td><td align="right"><strong>49.16</strong></td><td align="right">40.40</td><td align="right">30.73</td><td align="right">30.31</td><td align="right">32.87</td><td align="right"><u>45.91</u></td><td align="right"><strong>67.20</strong></td><td align="right"><u>54.42</u></td><td align="right">37.30</td><td align="right">52.14</td></tr>
-    <tr><td>CommerceAgentBench</td><td align="right"><u>37.38</u></td><td align="right">19.60</td><td align="right">9.30</td><td align="right">16.80</td><td align="right">30.80</td><td align="right"><strong>37.40</strong></td><td align="right"><strong>49.50</strong></td><td align="right"><u>46.30</u></td><td align="right">43.30</td><td align="right">39.30</td></tr>
-    <tr><td>Business Arena</td><td align="right"><strong>$79,868</strong></td><td align="right">$44,751</td><td align="right">$33,626</td><td align="right">$13,325</td><td align="right">$56,477</td><td align="right"><u>$66,292</u></td><td align="right"><strong>$168,867</strong></td><td align="right"><u>$89,423</u></td><td align="right">$40,804</td><td align="right">$55,742</td></tr>
-    <tr><td>GDPval<sup>†</sup></td><td align="right"><strong>1,128</strong></td><td align="right"><u>1,004</u></td><td align="right">869</td><td align="right">999</td><td align="right">951</td><td align="right">855</td><td align="right"><strong>1,741</strong></td><td align="right"><u>1,640</u></td><td align="right">1,500</td><td align="right">1,452</td></tr>
-    <tr><td>OfficeQA Pro</td><td align="right"><u>48.10</u></td><td align="right">39.10</td><td align="right">23.30</td><td align="right">46.60</td><td align="right">43.60</td><td align="right"><strong>59.40</strong></td><td align="right"><strong>74.40</strong></td><td align="right"><u>69.20</u></td><td align="right">51.20</td><td align="right">66.20</td></tr>
-    <tr><td>τ³-Bench (Banking)</td><td align="right"><strong>37.10</strong></td><td align="right">11.90</td><td align="right">7.20</td><td align="right"><u>25.80</u></td><td align="right">10.30</td><td align="right">21.70</td><td align="right"><u>46.90</u></td><td align="right"><strong>54.60</strong></td><td align="right">44.30</td><td align="right">37.10</td></tr>
-    <tr><td colspan="11" bgcolor="#DDF1EC"><font color="#087F6A"><strong>Tool calling</strong></font></td></tr>
-    <tr><td>AutomationBench (Pass¹)</td><td align="right"><strong>27.60</strong></td><td align="right">7.50</td><td align="right">2.20</td><td align="right">5.70</td><td align="right">14.80</td><td align="right"><u>18.50</u></td><td align="right"><strong>45.50</strong></td><td align="right"><u>43.50</u></td><td align="right">32.00</td><td align="right">28.00</td></tr>
-    <tr><td>AutomationBench (partial)</td><td align="right"><strong>69.10</strong></td><td align="right">39.40</td><td align="right">14.70</td><td align="right">27.90</td><td align="right">47.40</td><td align="right"><u>58.00</u></td><td align="right"><strong>81.20</strong></td><td align="right"><strong>81.20</strong></td><td align="right">59.70</td><td align="right"><u>70.00</u></td></tr>
-    <tr><td>BFCL v4</td><td align="right"><u>65.40</u></td><td align="right">63.19</td><td align="right">57.23</td><td align="right">62.81</td><td align="right">57.86</td><td align="right"><strong>68.51</strong></td><td align="right">64.33</td><td align="right"><strong>73.65</strong></td><td align="right">67.10</td><td align="right"><u>70.33</u></td></tr>
-    <tr><td>VitaBench</td><td align="right"><u>41.75</u></td><td align="right">34.25</td><td align="right">37.00</td><td align="right">26.25</td><td align="right"><strong>46.00</strong></td><td align="right">40.25</td><td align="right">46.75</td><td align="right"><u>52.25</u></td><td align="right"><strong>53.50</strong></td><td align="right">43.75</td></tr>
-    <tr><td colspan="11" bgcolor="#DDF1EC"><font color="#087F6A"><strong>Coding</strong></font></td></tr>
-    <tr><td>Terminal-Bench 2.1</td><td align="right">59.00</td><td align="right">49.50</td><td align="right">41.60</td><td align="right"><u>60.70<sup>*</sup></u></td><td align="right">33.70</td><td align="right"><strong>67.80<sup>*</sup></strong></td><td align="right"><strong>88.80</strong></td><td align="right">81.30<sup>*</sup></td><td align="right"><u>87.90<sup>*</sup></u></td><td align="right">82.70</td></tr>
-    <tr><td colspan="11" bgcolor="#DDF1EC"><font color="#087F6A"><strong>Instruction following</strong></font></td></tr>
-    <tr><td>IFEval</td><td align="right"><u>91.53</u></td><td align="right">86.90</td><td align="right"><strong>91.60</strong></td><td align="right"><strong>91.60</strong></td><td align="right">90.50</td><td align="right">81.80</td><td align="right"><u>95.00</u></td><td align="right"><strong>95.02</strong></td><td align="right">93.74</td><td align="right">93.89</td></tr>
-  </tbody>
-</table>
-
-Within each size group, **bold** denotes the best result and <u>underlining</u> denotes the second-best result. <sup>*</sup> Official model-card or Artificial Analysis result. <sup>†</sup> Reproduced on the public task release.
-
-### Cost-Performance
-
-<div align="center">
-  <picture>
-    <img src="assets/aggregate-cost-performance.png" width="100%" alt="Aggregate inference cost versus normalized performance across four co-work benchmarks">
+    <img src="https://raw.githubusercontent.com/Accio-Lab/occamy/9818009b54abbbc5f3c5620c93dfbf7e5b90dd06/assets/occamy-main-results.svg" width="100%" alt="Occamy-1.0 results on co-work, tool-use, coding, and business benchmarks">
   </picture>
 </div>
 
-Across Claw-Eval, WildClawBench, AutomationBench, and GDPval, Occamy-1.0 lies near the low-cost knee of the empirical Pareto frontier. Relative to its Qwen3.6-35B-A3B starting checkpoint, it delivers a large aggregate capability gain with only a modest change in measured per-task inference cost. Benchmark scores are equally weighted after per-benchmark min-max normalization, and costs are macro-averaged per task under the frozen pricing protocol used in the report.
-
+| Benchmark | Occamy-1.0 | Qwen3.6<br>35B-A3B | Agents-A1 | Nex-N2-mini | BigBang-1.0 | Ornith-1.5 |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Claw-Eval (average) | **82.2** | 69.5 | 69.9 | 66.6 | 63.5 | 64.4 |
+| Claw-Eval (Pass³) | **71.4** | 54.8 | 41.7 | 37.0 | 40.2 | 48.7 |
+| WildClawBench | **49.16** | 40.4 | 30.73 | 30.31 | 32.87 | 45.91 |
+| CommerceAgentBench | **37.40** | 19.6 | 9.3 | 16.8 | 30.8 | **37.40** |
+| Business Arena | **$79,868** | $44,751 | $33,626 | $13,325 | $56,477 | $66,292 |
+| GDPval | **1,128** | 1,004 | 869 | 999 | 951 | 855 |
+| OfficeQA Pro | 48.1 | 39.1 | 23.3 | 46.6 | 43.6 | **59.4** |
+| τ³-Bench (Banking) | **37.1** | 11.9 | 7.2 | 25.8 | 10.3 | 21.7 |
+| AutomationBench (Pass¹) | **27.6** | 7.5 | 2.2 | 5.7 | 14.8 | 18.5 |
+| AutomationBench (partial) | **69.1** | 39.4 | 14.7 | 27.9 | 47.4 | 58.0 |
+| BFCL v4 | 65.40 | 63.19 | 57.23 | 62.81 | 57.86 | **68.51** |
+| VitaBench | 41.75 | 34.25 | 37.00 | 26.25 | **46.00** | 40.25 |
+| Terminal-Bench 2.1 | 59.0 | 49.5 | 41.6 | 60.7<sup>*</sup> | 33.7 | **67.8<sup>*</sup>** |
+| IFEval | 91.53 | 86.90 | **91.60** | **91.60** | 90.50 | 81.80 |
+
+**Bold:** Best result in each row; ties are both bolded. <sup>*</sup> Official model-card result.
 
 ## 4. Training Recipe
 
-Occamy uses staged specialization and consolidation:
+Occamy is trained through two complementary specialization tracks, followed by parameter-space consolidation and a final SAO stage:
 
 ```text
-Qwen3.6-35B-A3B
-  ├─ Marathon Expert: SFT → HDPO ┐
-  └─ Sprint Expert: SFT          ├─ Uniform merge → SAO → Occamy-1.0
+                         Qwen3.6-35B-A3B
+                                  │
+                 ┌────────────────┴────────────────┐
+                 │                                 │
+                 ▼                                 ▼
+        Marathon Expert                     Sprint Expert
+           SFT → HDPO                             SFT
+                 │                                 │
+                 └────────────────┬────────────────┘
+                                  ▼
+                    Uniform parameter-space merge
+                                  │
+                                  ▼
+                                 SAO
+                                  │
+                                  ▼
+                             Occamy-1.0
 ```
 
 The Marathon Expert learns sustained execution and accuracy-conditioned efficiency, while the Sprint Expert preserves broader agentic capability. A uniform parameter-space merge combines both experts into one checkpoint with no inference-time routing or ensembling, and a final Single-Rollout Asynchronous Optimization (SAO) stage refines the merged policy on a broad co-work mixture.
@@ -247,6 +230,11 @@ This repository is released under the [Apache License 2.0](LICENSE). See the Hug
 
 ---
 
-## 8. Contact Us
+## 8. Contact and Correspondence
+
+For research inquiries, model questions, or collaboration opportunities, please contact the corresponding authors:
+
+- [Junbo Li](mailto:junboolee@gmail.com)
+- [Hongwei Xue](mailto:xuehongwe@gmail.com)
 
-For questions or feedback, please open an [issue](https://github.com/Accio-Lab/occamy/issues).
+For bug reports or feature requests, please open an [issue](https://github.com/Accio-Lab/occamy/issues).