Model card: current 7 models, per-model guidance from our testing, licensing and provenance

README.md

@@ -11,27 +11,56 @@ tags:
 
 The models OnBoard's installer and `ullama-recommend` offer, mirrored here for
 stable, provenance-clear hosting. All files are 4-bit GGUF quantizations for
-llama.cpp-based serving (shipped with OnBoard/Understand as `ullama`).
+llama.cpp-based serving (shipped with OnBoard/Understand as `ullama`), all
+under the Apache-2.0 license, mirrored unmodified from the source repos below.
+
+Every model here has been through our own qualification testing on real
+Understand project data: a chat test that measures how well the model answers
+questions about a codebase using OnBoard's analysis tools, and a code-summary
+benchmark that grades generated overviews for accuracy against the source.
+The notes below come from those measurements, not from vendor claims. That
+said, AI is AI — every model gets things wrong, so treat any answer as a
+starting point and take it with a grain of salt.
+
+## Which model should I use?
+
+Run `ullama-recommend` (shipped with OnBoard and Understand) — it looks at
+your machine's memory and suggests the right size. In short: pick the largest
+one your machine runs comfortably, and prefer the gemma models when you want
+the best chat answers.
+
+| Model | Download | Why you would pick it |
+|---|---|---|
+| gpt-oss-120b | 63 GB (2 parts) | The most accurate code summaries in our testing, and quick for its size. Needs roughly a 128 GB machine. |
+| Qwen3.6-35B-A3B | 22.4 GB | Highly accurate code summaries, and much faster than other models this large. |
+| Gemma 4 12B | 7.1 GB | The best chat answers of any model we tested. |
+| Qwen3.5 9B | 5.7 GB | Did well in chat testing, and digs a little deeper into code before answering. |
+| Gemma 4 E4B | 5.0 GB | Did well in our chat testing. |
+| Gemma 4 E2B | 3.1 GB | OnBoard's default model. Good at both chat and code summaries. |
+| Qwen3.5 2B | 1.3 GB | Writes good code summaries remarkably fast, but struggled in our chat testing. |
 
 Selection method, scores, and per-model sampling settings are documented in
 the [stinb/ullama](https://github.com/stinb/ullama) repo (`model_info.md`,
 `launcher/ullama-models.conf`).
 
+## Files, licensing, and provenance
+
 | File | Model | Company (Country) | License | Source |
 |---|---|---|---|---|
+| gpt-oss-120b-Q4_K_M-0000?-of-00002.gguf | gpt-oss-120b | OpenAI (United States) | Apache-2.0 | [unsloth/gpt-oss-120b-GGUF](https://huggingface.co/unsloth/gpt-oss-120b-GGUF) |
 | Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf | Qwen3.6-35B-A3B | Alibaba (China) | Apache-2.0 | [unsloth/Qwen3.6-35B-A3B-GGUF](https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF) |
-| gemma-4-31B-it-UD-Q4_K_XL.gguf | Gemma 4 31B | Google (United States) | Apache-2.0 | [unsloth/gemma-4-31B-it-GGUF](https://huggingface.co/unsloth/gemma-4-31B-it-GGUF) |
 | gemma-4-12b-it-Q4_K_M.gguf | Gemma 4 12B | Google (United States) | Apache-2.0 | [unsloth/gemma-4-12b-it-GGUF](https://huggingface.co/unsloth/gemma-4-12b-it-GGUF) |
-| Apriel-1.6-15b-Thinker-Q4_K_M.gguf | Apriel 1.6 15B Thinker | ServiceNow (United States) | MIT | [ServiceNow-AI/Apriel-1.6-15b-Thinker-GGUF](https://huggingface.co/ServiceNow-AI/Apriel-1.6-15b-Thinker-GGUF) |
 | Qwen3.5-9B-Q4_K_M.gguf | Qwen3.5 9B | Alibaba (China) | Apache-2.0 | [unsloth/Qwen3.5-9B-GGUF](https://huggingface.co/unsloth/Qwen3.5-9B-GGUF) |
 | gemma-4-E4B-it-Q4_K_M.gguf | Gemma 4 E4B | Google (United States) | Apache-2.0 | [unsloth/gemma-4-E4B-it-GGUF](https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF) |
-| Qwen3.5-4B-Q4_K_M.gguf | Qwen3.5 4B | Alibaba (China) | Apache-2.0 | [unsloth/Qwen3.5-4B-GGUF](https://huggingface.co/unsloth/Qwen3.5-4B-GGUF) |
 | gemma-4-E2B-it-Q4_K_M.gguf | Gemma 4 E2B | Google (United States) | Apache-2.0 | [unsloth/gemma-4-E2B-it-GGUF](https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF) |
 | Qwen3.5-2B-Q4_K_M.gguf | Qwen3.5 2B | Alibaba (China) | Apache-2.0 | [unsloth/Qwen3.5-2B-GGUF](https://huggingface.co/unsloth/Qwen3.5-2B-GGUF) |
-| Phi-4-mini-instruct-Q4_K_M.gguf | Phi-4 Mini | Microsoft (United States) | MIT | [unsloth/Phi-4-mini-instruct-GGUF](https://huggingface.co/unsloth/Phi-4-mini-instruct-GGUF) |
 
-Quantization policy: Q4_K_M for models ≤12B; Unsloth Dynamic (UD-Q4_K_XL)
-for the two large MoEs, where the dynamic quant's quality edge matters most.
+Quantization policy: Q4_K_M for models ≤12B; Unsloth Dynamic (UD-Q4_K_XL) for
+the large MoE, where the dynamic quant's quality edge matters most. gpt-oss-120b
+is a two-part split: download both parts into the same directory and point the
+server at part 1 — llama.cpp finds the second part automatically.
 
-All models retain their original licenses (Apache-2.0 / MIT, noted per file
-above); no modifications were made beyond mirroring the original GGUF files.
+All models retain their original Apache-2.0 licenses; no modifications were
+made beyond mirroring the original GGUF files. Models that failed our
+qualification testing (Qwen3.5 4B, Apriel 1.6, Gemma 4 31B, Phi-4 Mini) were
+removed from this repo on 2026-08-18.