[unrestricted-dev]

Adds Unrestricted (https://unrestricted.sh), an OpenAI-compatible API for uncensored models.

• qwen3.8-27b-uncensored: 262K context, $0.50 / $0.05 cached / $1.00 per 1M tokens
• deepseek-v4.1-flash-uncensored: 1M context, $0.35 / $0.10 / $1.25
• glm-5.3-flash-uncensored: 1M context, $0.20 / $0.19 / $0.60
• mimo-v2.6-flash-uncensored: 1M context, $0.40 / $0.29 / $1.30

All entries are schema v2. Each file lists its sources (our public /v1/models, /pricing and model pages); structured output and temperature were live-tested on our API for every model. bun validate passes.

[github-actions]

Action items

• [high] [possible mistake] providers/unrestricted/models/qwen3.8-27b-uncensored.toml:20 - Check: Reasoning options must copy the lab + same-surface peer set for that model, not invent GPT-style L/M/H. Why: Qwen3.8 peers use low/medium/xhigh with toggle/none+budget_tokens (providers/alibaba/models/qwen3.8-max.toml:6-18, providers/alibaba/models/qwen3.8-flash.toml:5-14, providers/deepinfra/models/Qwen/Qwen3.8-27B.toml:3-8, providers/openrouter/models/qwen/qwen3.8-27b.toml:1-11, providers/nebius/models/Qwen/Qwen3.8-27B.toml:8-25 explicitly notes high is an alias and lists xhigh), contradicting the PR's ["none","low","medium","high"] with no toggle/budget_tokens and high instead of xhigh; the PR's "first-party host" classification is unsupported for a Qwen-derived fine-tune served over an OpenAI-compatible API and a citation URL alone does not prove the disputed wire value. Action: Verify the exact wire controls on https://unrestricted.sh/reference and live /v1/chat/completions — whether off is effort=none vs separate enable_thinking toggle, whether graded levels are low/medium/xhigh vs high, and whether thinking_budget is exposed — then align to the lab/peer baseline (e.g. Nebius-style ["none","low","medium","xhigh"] if none truly disables, or toggle + ["low","medium","xhigh"] plus budget_tokens if those controls exist) with a leading wire comment; if high is truly canonical, provide docs/live-probe evidence that xhigh is rejected.

[unrestricted-dev]

Thanks, fixed in 1cff12040. The API's levels are none, low, medium and xhigh (high and max are accepted as aliases of xhigh, minimal of low), the same set as the Qwen3.8 27B peers, so reasoning_options is now [{ type = "effort", values = ["none", "low", "medium", "xhigh"] }] with a leading wire comment.

• none turns thinking off (live: no reasoning_content, 4 completion tokens), so there is no separate toggle.
• No reasoning budget field is documented on chat completions, so there is no budget_tokens option.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[unrestricted-dev]

Rebased on current dev and updated the entry to the current price list: $1.00 input / $0.10 cached input / $3.00 output per million tokens (https://unrestricted.sh/pricing). The description now matches what the API is mostly used for.

@rekram1-node this is ready for review whenever you have a moment. Happy to change anything.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [medium] [possible mistake] providers/unrestricted/models/deepseek-v4.1-flash-uncensored.toml:40 - Check: Provider structured_output/temperature must reflect this host's API, not peer copies. Why: Header admits both were copied without Unrestricted evidence and notes a peer split on structured_output (hopper/venice true vs nano-gpt false), so choosing true is arbitrary and may misstate json_schema support. Action: Live-probe json_schema enforcement and temperature on Unrestricted /v1/chat/completions as done for the qwen entry, or remove the claim until verified.
• [low] [possible mistake] providers/unrestricted/models/glm-5.3-flash-uncensored.toml:38 - Check: Provider capabilities need host evidence, not peer inheritance. Why: Header states structured_output = true/temperature = true were copied from nano-gpt/zai peers because Unrestricted does not state them, so unsupported values would be published as host facts. Action: Cite Unrestricted docs or add a live-test note for json_schema and temperature, or omit until verified.
• [low] [possible mistake] providers/unrestricted/models/mimo-v2.6-flash-uncensored.toml:37 - Check: Provider capabilities need host evidence, not single-peer copies. Why: Header states structured_output = false/temperature = true were copied from the sole uncensored peer while stock Xiaomi/Venice entries list true, so the host value is unverified and the choice is unexplained. Action: Cite Unrestricted docs or add a live-test note for json_schema and temperature, or omit until verified.

[github-actions]

No actionable findings.

[unrestricted-dev]

Updated: rebased on current dev, converted to schema v2, refreshed Qwen's pricing and added our three other live models, with structured output and temperature live-tested. bun validate passes. Ready for review. Thanks!

[github-actions]

No actionable findings.