LLM Profiles

Per-model tuning profiles that make AI agents call tools reliably across very different models. Tool-call reliability and the right sampling parameters vary a lot by model (first-attempt tool-call success ranges roughly Qwen ~94% > GLM ~91% > DeepSeek ~87%, and each family ships different recommended temperatures), so one global setting can never fit them all. A profile matches a provider/model and supplies the right knobs.

Built for Skales, usable as a plain data format anywhere.

What a profile tunes

• maxTools - cap the number of tools offered (helps models that choke on a big tool set).
• compactionLevel - prompt compaction floor (0 full, 1 compact, 2 minimal).
• promptHint - an instruction prepended to the system prompt: per-model guidance (decisiveness, grounding) plus the library's shared voice rules (act in the same turn you announce a tool, answer like a colleague, no filler or fake knowledge-cutoff disclaimers, check context before asking).
• params - sampling: temperature, top_p, top_k.
• toolCallStyle - native | json | xml: what the model does with tool calls. json/xml says it writes them as text, which switches on the host's fenced-block recovery so such a call still runs instead of showing up as the answer.
• toolHints - per-tool notes that teach a model your exact tool names (keyed by the real tool name), for models that reach for another framework's names like create_file or bash. Look the names up in TOOLS.md.

Frontier models (Claude, GPT, Gemini) intentionally have no profile here and run unchanged. Profiles target the models that benefit: weaker, local, or quirky.

Profiles in this library

All 28 profiles, grouped by family (index.json keeps them in the order they were added, which is the manifest's job, not a reading order). A profile with no sampling params carries the behaviour layer only: one tool at a time, no repeated call, exact tool names, say what a failed tool reported, and the system prompt outranks a squad brief. That is deliberate - this repo does not ship numbers it cannot cite.

Profile	Matches	Why
DeepSeek	deepseek	V3-era / R1 / local distills: low temp + explicit tool-name hints for reliable multi-tool calling.
DeepSeek V4	deepseek-v4	Agent-grade (V4 / Flash / Pro); vendor params (temp 1.0, top_p 1.0), no tool cap (it drives the full catalogue).
Qwen	qwen	Qwen2.5 / Qwen3: strong native tool-caller; vendor non-thinking params (temp 0.7, top_p 0.8, top_k 20). Also where qwen3-max lands, and correctly so.
Qwen 3.5	qwen3.5	Vendor params moved to temp 1.0, top_p 0.95, top_k 20.
Qwen 3.6	qwen3.6	qwen3.5 does not match 3.6, so it fell back to Qwen2.5-era params. Same vendor numbers as 3.5, plus the local dialect (a bare typed JSON object instead of a native call) named in the hint.
Qwen 3.7	qwen3.7	Thinking mode on every turn, so the generic qwen profile was handing it the vendor's NON-thinking numbers. Params are the vendor's own published 3.7 run config (temp 1.0, top_p 0.95, top_k 20), and longThinking keeps a long silent stretch from reading as a stalled stream.
MiniMax	minimax	Agentic tool-caller (M2.7 / M3); vendor params (temp 1.0, top_p 0.95, top_k 40), no tool cap (it drives the full catalogue).
MiniMax abab	abab	The older MiniMax namespace, including abab6.5s-chat, the provider's recommended pick: no abab id contains minimax, so the default model ran with no profile. No params - the M2.7/M3 cards say nothing about abab.
GLM	glm	Solid tool-caller (4.x / 5); GLM-5 agentic-eval params (temp 0.7, top_p 0.95).
GigaChat	gigachat	Sber's family: names tools that are not offered (web_search for search_web) and sometimes writes the call as XML in its text. Exact-name toolHints, native calls only, deterministic tool turns (toolParams temp 0.2).
Kimi	kimi	Over-deliberates; instant-mode params + hint to act decisively. Note the endpoint pins its temperature and rejects ours, so only top_p lands.
Mistral	mistral	Low vendor temp; good for structured output.
Magistral	magistral	Mistral's reasoning line. Its ids contain mistral, so a long-chain model was being handed Mistral Small's 0.15; its card says top_p 0.95, temp 0.7, word for word. The longer literal now wins.
Devstral	devstral	Mistral's 24B coding/agent model (matched no profile before); low deterministic temp + a coding-agent hint (edit precisely, verify with tests).
Codestral	codestral	codestral contains no mistral, so the code model matched nothing. No params - Mistral publishes none for it, and inheriting the family's 0.15 is the guess this repo does not make.
Mixtral	mixtral	Spelled with an x, so a bare Ollama mixtral matched nothing while mistralai/Mixtral-8x22B matched through the vendor prefix. No params, same reason as Codestral.
Gemma	gemma	Lower temp + compact prompt for reliable tool use (Gemma 2 era).
Gemma 4	gemma4	Measured, not inherited: Gemma 4 calls tools natively, and the card asks for temp 1.0, top_p 0.95, top_k 64 with a 131k window - the family profile's 0.3 and compaction floor were pure loss.
Gemma 4 (vendor id)	gemma-4	Google's own ids are hyphenated (google/gemma-4-27b-it), which gemma4 does not match. A deliberate twin, not a variant. The pattern is a literal, because a glob would also swallow gemma-3-4b-it.
Hunyuan Hy3	hy3	Tencent's flagship and the default Skales selects for the Hunyuan provider. Its ids contain no hunyuan substring at all. Sampling from the official vLLM recipe (temp 0.9, top_p 1.0).
Hunyuan (Tencent)	hunyuan	The other Hunyuan namespace on the legacy host (hunyuan-turbos, -vision, -a13b), which answers to neither hy3 nor anything else. No params - the A13B card recommends none.
Llama	llama	Vendor params.
Muse Glimmer	muse-glimmer	Meta's 30B agentic open-weights model (Apache-2.0, August 2026). Its ids carry no family substring and 30b is above the small-model bound, so it matched nothing. Vendor run config (temp 1.0, top_p 0.95, top_k 64), 131k window, and longThinking so its silent reasoning is not read as a stalled stream.
Nemotron	nemotron	NVIDIA Nemotron-3 (ultra / nano); middle-ground sampling, and longThinking so a long silent stretch is not read as a stalled stream.
Phi	phi	Verbose tool-caller; low temp + compact prompt.
GPT-OSS	gpt-oss	Strong native tool-caller for its class; decisive turns.
DBRX	dbrx	Databricks' open MoE instruct model. No family substring and no parameter count in the id, so it matched nothing and ran with no tool hints. No params - the 0.9/0.9 that circulates for it is a hosted playground default, not a vendor recommendation.
Small model	*b (<= ~9B ids)	Compact prompt for tiny models (tool set is trimmed by the host's local-model setting, not the profile).

How Skales uses them

In Skales: Settings → LLM Profiles (its own card, next to AI Providers) or the dedicated /profiles page (opt-in). The most specific match wins (longest literal pattern, provider pinned); an imported profile outranks a built-in at equal specificity, so you can override one. Import a profile by file, by pasting JSON, or by URL - point it at a raw file in this repo (from Skales v11.3.2 a normal GitHub file link works too; it is rewritten to the raw file automatically). index.json is a manifest so a client can list and fetch the current set and stay up to date as profiles are added.

Schema

See SCHEMA.md. A profile is a small JSON file; the format is kept deliberately simple so a profile is easy to write, review, and share.

Safety

A profile is data, not code, but promptHint and toolHints are injected into the model's system prompt, so an untrusted profile is a prompt-injection vector. Import only from sources you trust; Skales shows exactly what a profile injects before you rely on it, clamps params to sane ranges, and bounds the importer. See the Safety section in SCHEMA.md.

Contributing

Profiles are community-extensible. Add or improve one via PR, see CONTRIBUTING.md. Keep entries vendor-grounded (cite the source for params) and conservative (do not over-tune a capable model).

License

The profile data here is provided for use with Skales and compatible tools. Skales is BSL-1.1. These profiles are configuration data, not Skales source.