SLM Weekly #1 — what we added, what we skipped

[sha367]

SLM Weekly #1 — what we added, what we skipped

TL;DR: DeepSeek-R1-Distill-Qwen-1.5B (1.04 GB, Qwen2-architecture reasoning
model) is in the privateSLM catalog now. Qwen3.5-2B almost made it too. It
didn't: our vendored llama.cpp has no LLM_ARCH_QWEN3 entry, so the file
would not load. Full writeup with the actual grep output:
SLM Weekly #1
and the longer technical dive:
The number "60+ supported models" doesn't mean what you think.

Model	Size	Architecture	In catalog?
Llama 3.2 1B Instruct	808 MB	llama	yes
Qwen2.5 1.5B Instruct	1.0 GB	qwen2	yes
Gemma 2 2B Instruct	1.7 GB	gemma2	yes
DeepSeek R1 Distill Qwen 1.5B	1.04 GB	qwen2	yes (new)
Qwen3.5-2B	~1.2 GB	qwen3	no — engine gate

Open questions for the thread:

1. If you've run DeepSeek-R1-Distill-Qwen-1.5B locally (llama.cpp, LM Studio,
   Ollama, or otherwise) — what tokens/sec are you seeing, and on what
   device? Curious how it compares to the plain Qwen2.5-1.5B-Instruct we
   already ship.
2. We've traced the fix to llama.cpp PR #12828
   ("Support Qwen3 and Qwen3MoE"). If anyone has already rebased a vendored
   fork past that commit, did anything else break for you along the way?
3. For other local-AI app builders here: does your app publish (anywhere)
   the actual architecture list your vendored engine supports, or only the
   upstream project's full list? Genuinely curious how common this
   disclosure gap is.