h0ney-badger/qwen2.5-7b-universal-distill

qwen-uni-7b (v12) — a small, local universal model

Distilled/adapted from Qwen2.5-7B-Instruct into a single 7B that does coding + web design + general chat + tool/MCP function-calling — built on the code-distill pipeline, run entirely on one 16 GB GPU. GGUF: qwen-uni-7b-v12-Q5_K_M.gguf (fp16 weights also in this repo).

⚠️ REQUIRED: run it with a context window of at least ~9K tokens (16K recommended)

If you use this model with a tool/MCP server (e.g. searxng web search), set the context length to 9,216+ — 16,384 is a good default (the model supports 32,768).

Why this matters: an MCP server injects its tool schemas into the prompt (searxng's four tools are ~1,600 tokens), and a single web-search result is ~3,000 tokens. Together that is ~5,000 tokens of prompt before the model writes a word. In a 4K context those alone overflow the window, so the model has zero room to reply — which surfaces in LM Studio as "This message contains no content" and, if you nudge it, an endless re-search loop. This is a runtime configuration issue, not a model defect: at ≥9K (ideally 16K) it answers normally, and multi-turn searches don't choke. In LM Studio, set Context Length in the model's load config and reload.

The constant test (frozen suite, identical every version)

eval_universal.py scores C execution pass@1, HTML structural validity, tool-call schema-correctness (canonical <tool_call>, parsed via llama.cpp --jinja), and chat.

Headline: tool/MCP calling went 25% → 100% (correct tool, correct arguments, canonical <tool_call>), while coding/web/chat held. C sits in the 4–5/12 noise band every iteration showed — the documented small-model C ceiling.

What each iteration fixed (v6 → v12)

• —v6 — the core win: switched from the coder base to the general `Qwen2.5-7B-Instruct` base because the Coder base ships the tool tokens <tool_call>/</tool_call> with an all-zero (dead) embedding (Qwen ties embed→output, so a zero row can never be emitted). Tools 25% → 100%.
• —v8 — presentation: emits code in proper ``fenced blocks``, multi-line-formatted, and switches modes (artifact → plain follow-up gets prose).
• —v9/v10 — the artifact-mode tool-lock: once an artifact was in history the model would refuse to call a tool and fabricate; fixed with multi-turn enrich/research data that place a <tool_call> after an artifact.
• —v12 — after a large, messy search result, a plain "just tell me" ask sometimes got an HTML/markdown document instead of prose; retrained on realistic big-blob results with a prose-vs-doc intent discriminator. Measured in LM Studio on the real search blob: HTML-dump rate 15% → 2% at temp 0.8, other axes held.

How to use it

• —Tool-calling: serve with a current llama.cpp (llama-server --jinja) or a recent LM Studio runtime; older bundled runtimes mis-render the tool token. And set context ≥ 9K (see the warning above).
• —Coding / web / chat work in any runtime.
• —Apache-2.0 throughout (base, teacher = Qwen2.5-32B-Instruct, self-generated data).

Honest caveats

• —Context length is not optional with MCP tools — see the warning above.
• —It's a universal 7B, not a pure-C specialist: the old coder base scores ~1 noise-item higher on the hard C subset. For coding + tools + chat + web in one model, this is the pick.
• —Small-model ceiling on long multi-turn artifacts: deep in a tool conversation, asking it to expand and merge a lot of detail into one large HTML document, it tends to produce a compact doc and can repeat rather than fully incorporate everything. Single-shot "make a detailed page" requests are fine; iterative "put ALL of that inside the doc" is where the 7B shows its size.
• —Web score is structural validity, not visual-design taste.