feat: add PPIO provider

[hu-miao]

Adds the PPIO provider, serving 46 open-weight models through its OpenAI-compatible endpoint.

Provider

• api = "https://api.ppio.com/openai/v1", npm = "@ai-sdk/openai-compatible", env = ["PPIO_API_KEY"]
• Model ids mirror PPIO's own /openai/v1/models ids, so catalog keys match what the API accepts.

Models

46 open-weight models across DeepSeek, MiniMax, Moonshot AI, Qwen, Xiaomi, and Z.ai. Every entry uses base_model and overrides only what differs from the lab entry: cost, plus limit, modalities, structured_output, reasoning_options, and interleaved where PPIO's API reports something different.

Pricing

PPIO lists prices in CNY per million tokens. Converted at CNY→USD 6.718405 (2026-09-20, https://open.er-api.com/v6/latest/USD); the rate, date, and source are recorded in each model file's header.

Nine models use context-tiered pricing, expressed with [[cost.tiers]] and the lowest band as the base [cost].

Two models needed a note on where the rate applies:

• glm-4.7 — PPIO splits the 1–32000 context band by output length (≤200 output tokens at CNY 2/8, above at CNY 3/14). cost.tiers is context-only, so the general-case rate is used for that band. cache_write is published only for the 32000+ band.
• glm-4.6v — cache_read is published only for the 32768+ band.

Reasoning

reasoning_options follows PPIO's thinking-mode documentation:

Models	Options
deepseek-v4-pro-0813, deepseek-v4-flash-0731	toggle + effort (low/high/max)
glm-5.3, glm-5.3-flash, kimi-k3	effort (low/high/max) — thinking is always on
minimax-m3	toggle (thinking.type = adaptive/disabled)
remaining models	[] — no documented caller control on this host

For the models with []: PPIO's chat-completions reference exposes no general reasoning_effort or thinking parameter. Its enable_thinking is documented only for deepseek/deepseek-v3.2-exp and separate_reasoning only for deepseek/deepseek-r1-turbo, neither of which is in this PR.

interleaved = { field = "reasoning_content" } is set for kimi-k3, minimax-m3, minimax-m2.7, and minimax-m2.7-highspeed, which document that side channel (the MiniMax models via reasoning_split = true).

Validation

• bun validate passes
• bun run test in packages/sdk passes (23 tests)

[github-actions]

Action items

• [high] [possible mistake] providers/ppio/models/deepseek/deepseek-v4-pro.toml:5 - Check: Relay reasoning_options must match this host’s real controls (and must not use [] from uncertainty when lab/peers expose controls). Why: deepseek-v4-pro, deepseek-v4-flash, deepseek-v4.1-flash, and deepseek-v4-flash-vision-exp set reasoning_options = [], while the same PPIO catalog gives toggle + effort (low/high/max) to dated siblings deepseek-v4-pro-0813 / deepseek-v4-flash-0731, and first-party DeepSeek plus established relays expose thinking controls for these undated IDs. Empty means “no caller control,” not “unlisted in one doc page.” Action: Confirm on PPIO’s chat API whether these four IDs accept the same thinking.type / reasoning_effort surface as the dated snapshots (or first-party). If they do, author matching reasoning_options (and wire comments); only keep [] with host-specific evidence that these IDs truly reject caller control.
• [medium] [violation] providers/ppio/models/zhipuai/glm-4.5v.toml:15 - Check: After base_model, provider files must be override-only (no restated identical fields). Why: [modalities] input = ["text", "image", "video"] matches models/zhipuai/glm-4.5v.toml exactly, so the override is redundant and violates the override-only rule. Action: Remove the identical [modalities] block; keep only real deltas (e.g. limit.context = 65_536).
• [medium] [possible mistake] providers/ppio/models/deepseek/deepseek-v4-pro-0813.toml:6 - Check: Provider-authored interleaved when this host returns a reasoning side channel. Why: deepseek-v4-pro-0813 and deepseek-v4-flash-0731 enable thinking controls, but neither sets [interleaved]. First-party DeepSeek and most OpenAI-compatible peers for these models set field = "reasoning_content". Omitting it understates the response shape if PPIO splits reasoning the same way (as already documented for MiniMax via reasoning_split). Action: Verify PPIO chat completions for these IDs; if reasoning is returned on reasoning_content (or reasoning_details), add the matching [interleaved] block.

[github-actions]

Action items

• [high] [violation] providers/ppio/models/alibaba/qwen3.5-122b-a10b.toml:4 - Check: Every toggle must have a leading top-of-file wire comment. Why: Many PPIO models set reasoning_options with type = "toggle" but only have CNY/pricing headers—no Toggle: … path. AGENTS.md and the reasoning-options audit require the exact request field (e.g. thinking.type = enabled|disabled) above the first key so clients know the wire shape. Action: Add a leading # Toggle: thinking.type = enabled|disabled (or the measured values, e.g. adaptive|disabled for M3) on every toggle model that lacks it, including the Qwen 3.5/3.6/3.8-27b, DeepSeek V4 non-dated (deepseek-v4-flash, deepseek-v4-pro, deepseek-v4.1-flash, deepseek-v4-flash-vision-exp), Kimi (kimi-k2.6), Xiaomi MiMo, and Zhipu GLM toggle entries. Keep the existing comments on deepseek-v4-*-0813/-0731, minimax-m3, and kimi-k3.
• [medium] [possible mistake] providers/ppio/models/moonshotai/kimi-k3.toml:17 - Check: Provider limit overrides must reflect this host’s real caps, not lab context mistaken for output. Why: Lab moonshotai/kimi-k3 has limit.output = 131_072 and limit.context = 1_048_576. PPIO only overrides output = 1_048_576, which is the lab context size and is an order of magnitude above the lab output cap. Action: Confirm PPIO’s max completion tokens for kimi-k3 from pricing/API docs; if output is still ~131k, drop or correct the override; if context differs, override limit.context instead.
• [medium] [possible mistake] providers/ppio/models/alibaba/qwen3-235b-a22b-instruct-2507.toml:11 - Check: limit.output should be max completion tokens on this host. Why: Lab output is 16_384 with context 262_144; PPIO sets output = 260_000, nearly the full context window, which is atypical for completion limits and easy to confuse with context. Action: Verify against PPIO model/pricing docs and set the real max output (or context) accordingly.
• [medium] [possible mistake] providers/ppio/models/alibaba/qwen3.8-2.4t-a95b.toml:6 - Check: Relay reasoning_options must not be [] when the underlying model has documented caller controls this host may still expose. Why: Lab notes for this open twin document reasoning_effort (low/medium/xhigh). The PR only shows thinking.type is rejected; it does not establish that effort is unavailable. Empty options mean no control, not “toggle failed.” Action: Probe whether reasoning_effort is accepted and effective on PPIO; if yes, author the measured effort set (and wire comment); if no, keep [] and state that effort was tested and ignored/rejected.
• [low] [possible mistake] providers/ppio/provider.toml:12 - Check: Provider header comments should match authored model capabilities. Why: The note “Models not listed above expose no caller control” contradicts many model files (Qwen 3.5/3.6, GLM 4.x/5.x, MiMo, etc.) that set toggle. Action: Update the provider.toml summary so it lists the toggle-only cohort (or drop the blanket “not listed” claim).

[github-actions]

Action items

• [high] [possible mistake] providers/ppio/models/alibaba/qwen3.8-27b.toml:4 - Check: Relay reasoning_options must follow lab/same-surface peer controls for the model, not a stripped subset. Why: This entry is toggle-only, but same-host qwen3.8-2.4t-a95b and OpenRouter peers author effort low/medium/xhigh for the Qwen3.8 open-weight line. Dropping effort here understates caller control if PPIO accepts reasoning_effort. Action: Probe reasoning_effort on this model; if accepted, set reasoning_options to toggle + ["low", "medium", "xhigh"] (with an Effort wire comment); if rejected, document that explicitly in the header.
• [medium] [possible mistake] providers/ppio/models/deepseek/deepseek-v3.2.toml:4 - Check: On relays, [] means affirmative no caller control, not “side channel missing / not fully re-tested.” Why: Lab metadata describes hybrid thinking/non-thinking, and OpenRouter peers expose a toggle, but this file uses reasoning_options = [] while notes only that reasoning_content is absent. Missing side channel does not by itself prove toggle is unavailable. Action: Verify whether thinking.type = enabled|disabled turns reasoning on/off for this id; if it does, set { type = "toggle" } (and wire comment); keep [] only with an affirmative no-control note.
• [medium] [possible mistake] providers/ppio/models/moonshotai/kimi-k2.5.toml:4 - Check: Same relay [] rule vs lab/peer controls. Why: kimi-k2.5 is a reasoner; OpenRouter and other peers expose a toggle, while this file keeps [] and only notes no reasoning_content. Action: Confirm whether thinking.type toggles reasoning on PPIO; set toggle + [interleaved] if the side channel appears when enabled, or keep [] only with measured no-control evidence.
• [medium] [possible mistake] providers/ppio/models/moonshotai/kimi-k3.toml:17 - Check: Provider [limit] overrides must reflect this host’s real max completion, not a swapped context value. Why: Lab moonshotai/kimi-k3 has limit.output = 131_072 (context 1_048_576). This override sets output = 1_048_576, i.e. max output equal to full context, which is unusual and may be a context/output mix-up. Action: Confirm PPIO’s max completion tokens for kimi-k3; if the lab 131K (or another measured cap) is correct, fix limit.output.
• [low] [possible mistake] providers/ppio/models/zhipuai/glm-5.toml:18 - Check: Context overrides should match the host catalog, not a near-miss of the lab default. Why: Lab and most hosts use 204_800 for GLM-5; this sets 202_800 (a value seen on some GLM-5 turbo/v variants elsewhere). That is a one-digit-off delta without a stated PPIO source. Action: Verify PPIO’s published context for glm-5 and align to the measured value (likely 204_800 unless PPIO documents 202_800).

[github-actions]

Action items

• [high] [violation] providers/ppio/models/moonshotai/kimi-k2.6.toml:6 - Check: toggle must not pair with effort that already includes none. Why: Final options are toggle plus effort values that include none (and minimal). Per AGENTS.md / audit skill, when off is an effort value, author effort only — no toggle. Dual controls misrepresent the wire surface and break the none/toggle rule. Action: Drop toggle (and its wire comment), keep effort with the real graded set including none (or document a separate on/off field that is not redundant with none).
• [high] [violation] providers/ppio/models/xiaomi/mimo-v2.5.toml:6 - Check: toggle must not pair with effort that already includes none. Why: Same dual shape: toggle + ["none", "low", "medium", "high"]. Patch 4 states none turns thinking off, so off is already covered by effort. Action: Use effort-only (with none); remove toggle.
• [high] [violation] providers/ppio/models/xiaomi/mimo-v2.5-pro.toml:6 - Check: toggle must not pair with effort that already includes none. Why: toggle + full enum including none/minimal. Same rule break as above. Action: Effort-only with verified values; no toggle.
• [high] [violation] providers/ppio/models/zhipuai/glm-5.1.toml:6 - Check: toggle must not pair with effort that already includes none. Why: toggle + ["none", "minimal", "low", "medium", "high", "xhigh", "max"]. Action: Effort-only (or toggle-only if effort is not a real graded control); never both when none is in effort.
• [high] [violation] providers/ppio/models/zhipuai/glm-5.2.toml:6 - Check: toggle must not pair with effort that already includes none; baseline = lab effective set. Why: Same dual none+toggle shape. First-party Zhipu/Z.ai GLM-5.2 is effort high|max (aliases map into those); dumping the full schema enum plus toggle invents a control surface the lab does not publish. Action: Remove toggle when none is present; author the effective levels this host actually distinguishes (likely none + high/max, or lab-like high/max if off is not real), not the full enum.
• [high] [possible mistake] providers/ppio/models/alibaba/qwen3.8-2.4t-a95b.toml:5 - Check: Relays must not use reasoning_options = [] when lab/same-surface peers expose caller control. Why: Lab and OpenRouter peers use always-on thinking with reasoning_effort low|medium|xhigh. Final file sets [] while the header still claims that effort set. Rejecting thinking.type only rules out toggle, not effort. Empty means “no caller control,” not “toggle unavailable.” Action: Restore effort ["low", "medium", "xhigh"] if the host accepts it, or prove effort has no effect and remove the effort comment; do not leave [] with a native-effort claim.
• [medium] [possible mistake] providers/ppio/models/moonshotai/kimi-k2.5.toml:5 - Check: Effort values must be the host’s real set, not a full-schema dump; baseline is lab/peer toggle-only for K2.5. Why: Options list none|minimal|low|medium|high|xhigh. Lab/established peers are mostly toggle-only. Patch 4 only claims none/minimal off vs low+ on — not that every graded level is distinct. Action: Keep effort-only (shape OK) but shrink values to levels with a meaningful, measured difference (e.g. off vs on, or documented graded set); avoid dumping unused schema tokens.
• [medium] [possible mistake] providers/ppio/models/deepseek/deepseek-v3.2.toml:5 - Check: [] only when there is affirmative no caller control on this host. Why: Header notes reasoning_effort is accepted but no side channel appears. That does not establish that thinking.type (or another on/off control) is unavailable on a hybrid-reasoning model. Action: Probe toggle (and any real off path); if thinking can be disabled, author toggle (and effort only if graded levels matter); keep [] only if controls truly do nothing.

[github-actions]

Action items

• [high] [violation] providers/ppio/models/zhipuai/glm-5.1.toml:4 - Check: reasoning_options must match this host’s real controls and must not dump non-distinct effort aliases (AGENTS.md / audit skill: baseline = lab + meaningful effect; never dump the full effort enum). Why: The PR’s own probe says for glm-5.1/5.2, kimi-k2.5/2.6, and mimo-v2.5/pro, none turns thinking off and every graded value turns it on with no ordered difference—yet these files still advertise long ladders (none/low/medium/high/xhigh/max). That invents graded caller control the host does not expose and diverges from lab baselines (glm-5.1/mimo/kimi-k2.x: toggle; glm-5.2: effective high/max). Action: Collapse each of glm-5.1, glm-5.2, kimi-k2.5, kimi-k2.6, mimo-v2.5, and mimo-v2.5-pro to the effective surface only (e.g. lab-style toggle, or effort with none plus the levels that actually change behavior—for glm-5.2 typically none/high/max), and align the leading effort comments with the values you keep.
• [medium] [possible mistake] providers/ppio/models/deepseek/deepseek-v4-pro.toml:7 - Check: DeepSeek V4 Pro effort baseline is lab/peer-effective levels, not every accepted alias. Why: First-party DeepSeek Pro is high/max (low/medium map to high). These PPIO Pro entries list low/high/max as if low were a distinct grade; if PPIO only accepts low as an alias of high, that overstates control. Action: Verify whether reasoning_effort=low is behaviorally distinct on this host for deepseek-v4-pro and deepseek-v4-pro-0813; if not, drop low and match the lab set high/max (keep Flash’s low/high/max if Flash still treats low as real).
• [medium] [possible mistake] providers/ppio/models/moonshotai/kimi-k3.toml:17 - Check: Provider [limit].output overrides must reflect this host’s real max output, not context size. Why: Lab moonshotai/kimi-k3 has output = 131_072 and context = 1_048_576; this entry sets output = 1_048_576, which looks like context copied into output and would mislead clients on max completion length. Action: Confirm PPIO’s documented/API max output for kimi-k3; if it is not 1M completion tokens, set limit.output to the real cap (or omit if it matches the lab 131_072).

[github-actions]

Action items

• [high] [possible mistake] providers/ppio/models/alibaba/qwen3-235b-a22b-instruct-2507.toml:11 - Check: Provider limit overrides must reflect this host’s real context/output caps, not a mislabeled peer value. Why: Lab metadata has limit.output = 16_384 and limit.context = 262_144. This file only sets output = 260_000, which is far above the lab completion cap and matches values other hosts use as context (e.g. scaleway context = 260_000), not max output. That would publish a false 260k completion window. Action: Verify PPIO’s reported max_tokens / context for this model id; if 260k is context, set [limit].context (and keep a real output cap); if output truly differs from the lab, cite the host figure and keep only the real delta.
• [low] [possible mistake] providers/ppio/provider.toml:15 - Check: Provider header comments must match the final measured model metadata. Why: The comment still says deepseek-v3.2 and kimi-k2.5 return no reasoning side channel, but later commits set both to reasoning_options with toggle and [interleaved] field = "reasoning_content". Stale host notes will mislead the next edit. Action: Update the provider.toml reasoning summary so it matches the final per-model files (side channel + toggle cohorts).

[github-actions]

No actionable findings.