Model Cards

Intended use

Autonomous and semi-autonomous code editing, refactors, and shell-level tasks issued through openclaw-agent. Runs alongside a native Ollama provider path that was left untouched — this is an additive LiteLLM route, not a replacement.

Limitations

• Quantized weights — no benchmark run against the full-precision checkpoint on this fleet; treat generated code as unreviewed until tested, same as any LLM output.
• No sandboxed execution boundary beyond openclaw-agent's own allowlist — the model does not independently rate-limit or scope its own tool calls.

Deployment notes

Version pins exist because litellm==1.95.0 breaks on newer FastAPI (get_flat_dependant import error) and langfuse 4.x drops the attribute LiteLLM's callback expects. Upgrade either only after re-verifying both constraints.

Intended use

Fast triage and classification ahead of a heavier model or a human decision — e.g. sorting inbound email into Work / Personal / Tour / Other / Spam and producing a structured {category, needs_reply, draft_reply} object for review.

Responsible AI design choices

• Draft-only, human-in-the-loop: the email classifier chatflow stops at a JSON draft. No send/label action is wired to it — a person reviews and sends, by design, not by omission.
• Local-only by policy: an earlier OpenAI-backed variant of this same flow was retired in favor of qwen3:8b specifically to keep the pipeline at zero cloud calls, rather than carry a design that broke that guarantee.

Limitations

Selected in part because the flow's original router (qwen2.5:7b) aged out of the registry — re-validate classification accuracy after any future model swap rather than assuming continuity.

Intended use

Backs every retrieval-augmented chatflow in the fleet — documents and notes are embedded once at ingest and queried at chat time against Qdrant, rather than re-embedded per request.

Limitations

• Retrieval quality has not been measured with a held-out eval set — there is no golden-set or regression suite behind these chatflows yet, so silent retrieval drift after a document-set change would currently go unnoticed.
• Single embedding model across all collections — no per-domain fine-tuning, so domain-specific jargon retrieves at whatever quality the base model provides.

Intended use

Default conversational model surfaced to Open WebUI users — general Q&A, drafting, and reasoning where no specialized router has already selected a narrower model.

Limitations

No model-selection telemetry exists yet to confirm this is actually the model most sessions land on — Open WebUI allows per-session override, and usage hasn't been audited against Langfuse traces to date. Treat this card's "default" framing as intent, not a measured fact, until that's checked.

Intended use

Text-to-image generation and instruction-driven edits, including multi-image reference chaining (e.g. transferring a garment from one reference image onto a subject in another).

Limitations

• No content-safety filter layered in front of this endpoint at the application level — anything ComfyUI itself doesn't refuse will render.
• Multi-image compositing (garment transfer, subject blending) has an unreviewed likeness/consent surface — treat outputs involving real people's photos with the same care as any other image edit of a real person, on or off this fleet.

Intended use

Grammar, spelling, and mechanics review of a saved or pasted resume — feedback only, does not rewrite structure or content.

Responsible AI design choices

Decoding limits above exist because of a real incident: the original config (repeat_penalty 1 — no penalty — with no num_predict cap and Qwen3's thinking mode never disabled) let one request run past 50,000 tokens without stopping. Fixed with the caps shown here plus /no_think appended to the outgoing prompt.

Limitations

Session+candidate-keyed caching means a substantive resume edit followed by an immediate re-ask can replay a stale cached result rather than regenerate — an accepted trade-off against the alternative (redundant-call storms that stalled real requests).

Intended use

Restructures and cleans up a resume's formatting. Explicitly forbidden from cutting content — shortening is condense-agent's job, kept as a separate model so "reformat" can never silently lose material.

Responsible AI design choices

Caught fabricating content in production — invented an Executive Summary section, fake percentages, and invented months/cities for date ranges the source never specified. Fixed with explicit verbatim-grounding rules in the Modelfile, verified clean against real resumes afterward. Also affected by a silent-truncation mode on long resumes (12KB+): the tool now fetches a saved candidate's text directly from docgen instead of asking the model to re-type it as a JSON argument.

Limitations

The p90 here is a known Ollama daemon-hang signature (multiple requests completing at an identical timestamp), not generic slowness — mitigated by a watchdog service since 2026-08-11, whose real catch-and-recover behavior against a live hang is still unverified in production.

Intended use

Shortens a resume to a target length by cutting and tightening the weakest/oldest material first. Every output ends with a CUT SUMMARY: line naming everything removed, so nothing is dropped silently.

Responsible AI design choices

Prompt-only "don't invent sections" rules didn't hold: a live test produced a fabricated "Technical Skills" section listing "Linux" — a skill absent from the actual source resume. Fixed with a code-level check that strips any output heading whose exact text doesn't appear in the source, with follow-up fixes for whitespace artifacts (e.g. a PDF-extracted no-break space) and legitimate heading merges (e.g. "Education" + "Professional Development" → "Education & Certifications") so the guardrail doesn't flag real edits as fabrication.

Limitations

Early version cached by content fingerprint, but the orchestrator re-serializes long text slightly differently each call — so a single turn could trigger 7+ live model calls before a cache hit ever landed. Fixed by dropping the fingerprint and keying on session+candidate+target only.

Intended use

Qualitative feedback on length, quantification, targeting, and narrative strength — critique only, produces no rewritten output of its own.

Limitations

Same Ollama-hang latency signature as the other specialist agents; shares GPU scheduling with the fleet's general chat model, so real-world response time depends on what else is loaded at request time, not just this model's own cost.

Intended use

Drafts a cover letter from a candidate's saved resume against one specific job posting.

Responsible AI design choices

Code-level guardrail strips any generated sentence that claims a technology or term from the job posting which the candidate's own resume never mentions — a fabrication check applied to the actual output, not left to prompt instructions alone.

Intended use

Produces a justified 0–100 fit score against one specific job posting, with named strengths, gaps, and a self-check of its own reasoning.

Limitations

The score is model-justified, not validated against actual hiring outcomes — treat it as a structured opinion to inform a decision, not a calibrated probability.