ADFA-6321 | Add Qwen2.5-Coder 1.5B Instruct to Get AI Models

plugins/Get-AI-Models/README.md

@@ -140,8 +140,9 @@ explicitly, waiving the instruction-tuned, open-licence and 16k-context gates fo
 so `CatalogLoader` no longer enforces the licence or context gates, and the unit tests no longer
 assert them. `docs/CURATION.md` records exactly which gate each entry breaks.
 
-Version 1.0.0 ships 6 files, all under 1 GB: Qwen2.5 0.5B (base), Qwen3 0.6B, Qwen3.5 0.8B,
-Llama 3.2 1B Instruct, SmolLM2 360M Instruct, and H2O Danube3 500M (base). Five are Apache-2.0;
+The catalog lists 7 files: Qwen2.5 0.5B (base), Qwen3 0.6B, Qwen3.5 0.8B, Llama 3.2 1B Instruct,
+SmolLM2 360M Instruct, H2O Danube3 500M (base), all under 1 GB, and Qwen2.5-Coder 1.5B Instruct
+(1.04 GiB), the only code-trained model. Six are Apache-2.0;
 Llama 3.2 1B is under the Llama 3.2 Community Licence.
 
 **The two behavioural gates are not yet proven.** Per the ticket, the admission harness that runs a

plugins/Get-AI-Models/docs/CURATION.md

@@ -58,9 +58,9 @@ All 12 entries passed that final HEAD cross-check, re-run after the most recent
 
 ## Catalog entries (maintainer-selected)
 
-**This catalog is a maintainer override, not a gate-derived list.** The six models below were named
-explicitly by the ticket owner, replacing the eleven gate-passing entries that shipped earlier. Three
-of the six break at least one hard gate; those gates are therefore **no longer enforced** in
+**This catalog is a maintainer override, not a gate-derived list.** The first six models below were
+named explicitly by the ticket owner, replacing the eleven gate-passing entries that shipped
+earlier; the seventh was requested later in ADFA-6321. Three of the six break at least one hard gate; those gates are therefore **no longer enforced** in
 `CatalogLoader` (licence, 16k context) nor asserted in `CatalogLoaderTest`.
 
 | Entry id | Base model | GGUF uploader | Ctx | Licence | Gate waived |
@@ -71,8 +71,14 @@ of the six break at least one hard gate; those gates are therefore **no longer e
 | `llama-3-2-1b-instruct-q4_k_m` | `meta-llama/Llama-3.2-1B-Instruct` | unsloth | 131072 | llama3.2 | **2** (Llama 3.2 Community Licence; upstream repo gated) |
 | `smollm2-360m-instruct-q8_0` | `HuggingFaceTB/SmolLM2-360M-Instruct` | Hugging Face TB | 8192 | apache-2.0 | **7** (8192 context) |
 | `h2o-danube3-500m-base-q4_k_m` | `h2oai/h2o-danube3-500m-base` | mradermacher | 8192 | apache-2.0 | **1** (base model), **7** (8192 context) |
+| `qwen2-5-coder-1-5b-instruct-q4_k_m` | `Qwen/Qwen2.5-Coder-1.5B-Instruct` | Qwen (Alibaba Cloud) | 32768 | apache-2.0 | none |
 
-Gates 3 and 4 remain unproven for all six, as before. Gate 5 (chat template) does not apply to the
+`qwen2-5-coder-1-5b-instruct-q4_k_m` was added by ADFA-6321: it is the only code-trained model in
+the catalog. Its gates 1, 2, 5, 6 and 7 were re-checked against the API on 2026-10-02, and
+`minRamBytes` is 4 GiB because it is the largest file here (1.04 GiB), one step above the 3 GiB of
+the ~0.8 GB entries.
+
+Gates 3 and 4 remain unproven for all seven, as before. Gate 5 (chat template) does not apply to the
 two base models - they have none, which is part of why they are not usable as assistants. Gate 6
 holds for every entry: single file, sha256 pinned to a commit revision, HEAD-verified.
 

plugins/Get-AI-Models/get-ai-models.html

@@ -128,10 +128,19 @@ <h3>The catalog</h3>
         <td>apache-2.0</td>
         <td>mradermacher</td>
     </tr>
+    <tr>
+        <td>Qwen2.5-Coder 1.5B Instruct</td>
+        <td><code>Q4_K_M</code></td>
+        <td>1.5B</td>
+        <td>1.04 GB</td>
+        <td>32k</td>
+        <td>apache-2.0</td>
+        <td>Qwen (Alibaba Cloud)</td>
+    </tr>
 </table>
 <p>Quantization is deliberately not restricted: <code>Q4_K_M</code> and <code>Q8_0</code> both appear
 because the gates are about behaviour, not about a preferred format.</p>
-<p><b>This list is a maintainer override, not a gate-derived set.</b> Five entries are Apache-2.0;
+<p><b>This list is a maintainer override, not a gate-derived set.</b> Six entries are Apache-2.0;
 Llama 3.2 1B is under the Llama 3.2 Community Licence, whose upstream repo is also access-gated. Two
 entries are pretrained <i>base</i> models with no chat template, and two have an 8k context window.
 The licence and context gates are consequently no longer enforced in code - see

plugins/Get-AI-Models/src/main/assets/catalog/models.json

@@ -102,6 +102,23 @@
       "baseModel": "h2oai/h2o-danube3-500m-base",
       "description": "H2O's small pretrained model, Apache-2.0. A base model, NOT instruction-tuned, so it continues text instead of answering, and it has an 8k context window. Strengths: very small, permissively licensed, trained on a well-described corpus. Weaknesses: unusable as an assistant as shipped, no chat template, no tool use, and below the 16k context floor. Listed at the ticket owner's request; it meets neither the instruction-tuned nor the context gate.",
       "behaviouralGatesVerified": false
+    },
+    {
+      "id": "qwen2-5-coder-1-5b-instruct-q4_k_m",
+      "name": "Qwen2.5-Coder 1.5B Instruct",
+      "quantization": "Q4_K_M",
+      "parameters": "1.5B",
+      "fileName": "qwen2.5-coder-1.5b-instruct-q4_k_m.gguf",
+      "sizeBytes": 1117320768,
+      "sha256": "cc324af070c2ecbfd324a30884d2f951a7ff756aba85cb811a6ec436933bb046",
+      "url": "https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF/resolve/f86cb2c1fa58255f8052cc32aeede1b7482d4361/qwen2.5-coder-1.5b-instruct-q4_k_m.gguf",
+      "minRamBytes": 4294967296,
+      "publisher": "Qwen (Alibaba Cloud) - official GGUF release",
+      "contextTokens": 32768,
+      "license": "apache-2.0",
+      "baseModel": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+      "description": "The only model here trained specifically on code, and the one to pick for help with Kotlin and Java. Strengths: noticeably better at code syntax and less prone to inventing APIs than the general chat models in this catalog, a 32k context window, a chat template with the Qwen tool-call format, and Apache-2.0 on both the model and the official GGUF. Weaknesses: the largest download here at about 1 GB, and it needs more memory than the sub-1B models, so it is slower on a mid-range phone; at 1.5B parameters with Q4_K_M rounding it still gets multi-step reasoning and less common libraries wrong, and it is weaker at general conversation than at code.",
+      "behaviouralGatesVerified": false
     }
   ]
 }