feat: update IteraCompute model catalog

[K-3-LT]

Summary

• expand IteraCompute from 2 to 9 live production models
• move the existing Qwen and Ornith entries to the canonical IDs returned by GET https://api.iteracompute.com/v1/models
• add current pricing, cache-read pricing, context/output limits, modalities, and provider-specific reasoning controls

Models

• qwen/qwen3.8-27b
• ornith-ai/ornith-1.5-35b-a3b
• qwen/qwen3.8-2.4t-a95b
• deepseek/deepseek-v4-flash-0731
• deepseek/deepseek-v4-pro-0813
• z-ai/glm-5.3
• z-ai/glm-5.3-flash
• minimax/minimax-m3
• moonshotai/kimi-k3

Validation

• git diff --cached --check
• �un validate

Source of truth: https://api.iteracompute.com/v1/models

[github-actions]

Action items

• [high] [violation] providers/iteracompute/models/deepseek/deepseek-v4-flash-0731.toml:7 - Check: Relay reasoning_options must follow lab + same-surface peers; [] only when this host truly has no caller control. Why: IteraCompute is a multi-model OpenAI-compatible relay. First-party DeepSeek V4 Flash and established peers use toggle + effort (e.g. low/high/max or high/max), and this same provider already authors reasoning_effort controls for Qwen/Ornith. Catalog silence is not affirmative proof of no control. Action: Replace [] with the lab/peer baseline this host actually forwards (typically toggle + DeepSeek effort), or keep [] only with host-specific docs/live proof that reasoning controls are rejected/ignored—not only GET /v1/models.
• [high] [violation] providers/iteracompute/models/deepseek/deepseek-v4-pro-0813.toml:7 - Check: Same relay reasoning baseline rule for DeepSeek V4 Pro. Why: Lab entry is toggle + high/max (Fireworks peers match; OpenRouter also exposes controls). [] erases known caller controls on a relay. Action: Author lab-aligned options (toggle + high/max, with a leading toggle wire comment) unless this host is proven not to forward them.
• [high] [violation] providers/iteracompute/models/minimax/minimax-m3.toml:7 - Check: MiniMax-M3 relay options vs first-party MiniMax. Why: First-party providers/minimax/models/MiniMax-M3.toml and faithful peers use a thinking toggle; [] is reserved for verified always-on/no-control serves, not catalog omission. Action: Use lab/peer toggle (with leading wire comment) unless IteraCompute is shown not to accept the MiniMax thinking switch.
• [high] [violation] providers/iteracompute/models/moonshotai/kimi-k3.toml:8 - Check: Kimi K3 relay options vs Moonshot lab. Why: Lab is toggle + effort low/high/max; peers copy that. [] contradicts the required relay baseline. Action: Copy the lab option set (and wire comment) unless this host is proven not to expose those controls.
• [high] [violation] providers/iteracompute/models/z-ai/glm-5.3.toml:8 - Check: GLM-5.3 relay options vs Z.AI lab. Why: Lab/peers use mandatory reasoning with effort low/high/max (this PR already does that correctly for glm-5.3-flash). [] is inconsistent and drops graded control. Action: Set reasoning_options to effort low/high/max like the lab and flash sibling, unless host-specific evidence shows effort is unavailable.
• [high] [violation] providers/iteracompute/models/qwen/qwen3.8-2.4t-a95b.toml:7 - Check: Qwen3.8 2.4T A95B option shape must match lab/peer baseline; every toggle needs a leading wire comment. Why: Lab and peers (OpenRouter, DeepInfra, Fireworks) treat thinking as always-on with effort low/medium/xhigh, not a bare toggle. A lone toggle invents a different control surface and lacks the required top-of-file wire path. Action: Use effort ["low", "medium", "xhigh"] (no toggle) unless this host uniquely exposes only on/off; if toggle stays, document the exact request field in a leading comment.
• [medium] [possible mistake] providers/iteracompute/models/ornith-ai/ornith-1.5-35b-a3b.toml:6 - Check: Provider overrides must retain real host deltas after ID moves. Why: The deleted entry overrode attachment = false and [modalities] input = ["text"] for a language-only NVFP4 serve. The new file drops both, so it inherits lab attachment = true and image input. Action: Restore the text-only/attachment overrides if still accurate, or cite host evidence that this deployment now accepts images.
• [medium] [possible mistake] providers/iteracompute/models/qwen/qwen3.8-2.4t-a95b.toml:6 - Check: Override-only fields must be true host deltas. Why: Lab metadata is text-only (attachment = false, modalities text). This sets attachment = true and image input without a mapped host claim beyond the generic models API URL. Action: Confirm IteraCompute’s qwen/qwen3.8-2.4t-a95b accepts images; if not, drop attachment and the image modality override.
• [low] [possible mistake] providers/iteracompute/models/moonshotai/kimi-k3.toml:17 - Check: limit.output should be the real host max completion budget. Why: output = 999_999 looks like a catalog sentinel versus lab 131_072 and typical peer values. Action: Verify against the provider’s documented/API max completion tokens and replace the sentinel if it is not a real limit.

[K-3-LT]

@rekram1-node Hi, could you please take a look at this PR when you have a chance? All 2 checks are passing and it has no merge conflicts. The automated review also raised several catalog metadata action items; if any of them are blocking the merge, we'd appreciate guidance on the expected changes. Thanks!

[K-3-LT]

@rekram1-node Hi, we have updated this PR to match the current production pricing returned by GET /v1/models (commit ffc5b282e). The changed rates, in USD per million input / cached-input / output tokens, are:

• DeepSeek V4 Flash: $0.34 / $0.035 / $1.05
• DeepSeek V4 Pro: $1.10 / $0.11 / $3.30
• MiniMax M3: $0.29 / $0.08 / $1.20
• Kimi K3: $3.00 / $0.29 / $14.90
• GLM 5.3 Flash: $0.14 / $0.03 / $0.49
• GLM 5.3: $1.20 / $0.26 / $3.50

The other three model prices are unchanged. Could you please review and merge when the checks pass? We also saw the automated metadata review; if any of those items are merge blockers, please let us know which host-specific evidence or changes you need. Thanks!

[github-actions]

Action items

• [high] [violation] providers/iteracompute/models/deepseek/deepseek-v4-flash-0731.toml:7 - Check: Relay reasoning_options must follow lab/same-surface peer controls; [] only when this host truly has no caller control. Why: IteraCompute is a multi-model OpenAI-compatible relay. First-party DeepSeek and established peers for this snapshot expose toggle and/or effort (e.g. high/max), but this entry sets reasoning_options = [] based only on the public models catalog not listing controls. Empty means “no control,” not “catalog didn’t document it,” so callers lose real reasoning controls. Action: Replace [] with the controls this host actually forwards (baseline: lab/peers toggle + high/max, or the host’s verified effort set), and document the exact wire paths in a leading comment.
• [high] [violation] providers/iteracompute/models/deepseek/deepseek-v4-pro-0813.toml:7 - Check: Same relay reasoning baseline for DeepSeek V4 Pro. Why: Lab providers/deepseek/models/deepseek-v4-pro.toml and peers use toggle + high/max (or equivalent). reasoning_options = [] with a “catalog exposes no control” rationale is the uncertainty anti-pattern for relays. Action: Author the host’s real DeepSeek V4 Pro controls (not []) and cite the request fields.
• [high] [violation] providers/iteracompute/models/moonshotai/kimi-k3.toml:10 - Check: Kimi K3 relay options must match lab/peer surface. Why: Lab and peers use toggle plus graded effort (low/high/max or subset). This file uses [] while claiming reasoning is mandatory with no control, which contradicts the lab API and peer relays. Action: Set reasoning_options to the controls this host forwards (at minimum lab-shaped toggle + effort if passthrough), with leading wire-path comments.
• [high] [violation] providers/iteracompute/models/z-ai/glm-5.3.toml:8 - Check: GLM-5.3 always-on thinking still has effort controls. Why: First-party Zhipu/Z.AI and peers author effort low/high/max (thinking not disableable, but effort is selectable). [] incorrectly implies no caller control and diverges from the lab baseline this host should copy unless proven different. Action: Use [{ type = "effort", values = ["low", "high", "max"] }] (or this host’s verified equivalent), matching glm-5.3-flash on the same provider.
• [high] [violation] providers/iteracompute/models/qwen/qwen3.8-2.4t-a95b.toml:7 - Check: Qwen3.8 2.4T-A95B baseline is effort, not bare toggle; every toggle needs a leading wire comment. Why: Lab metadata and major peers (OpenRouter, Cortecs, Kilo, DeepInfra) use always-on reasoning with effort low/medium/xhigh. A lone { type = "toggle" } invents a binary control those surfaces do not use, omits the graded efforts, and has no top-of-file wire path. Action: Replace with the peer/lab effort set (and only add toggle if this host documents a separate on/off field), with a leading wire comment for any toggle.
• [medium] [violation] providers/iteracompute/models/minimax/minimax-m3.toml:8 - Check: MiniMax-M3 first-party exposes a reasoning toggle. Why: providers/minimax/models/MiniMax-M3.toml authors toggle; several relays keep a toggle. [] is only valid with affirmative proof this host has no control, not from catalog silence. Action: Verify IteraCompute’s request surface; if it forwards MiniMax thinking on/off, author toggle (with wire comment); keep [] only with host-doc or live proof of no control.
• [medium] [possible mistake] providers/iteracompute/models/ornith-ai/ornith-1.5-35b-a3b.toml:6 - Check: Provider overrides must retain real host deltas vs lab. Why: The deleted entry forced language-only serving (attachment = false, modalities.input = ["text"]) for the NVFP4 build. The replacement drops those overrides, so merge inherits lab attachment = true and image input, which may mis-describe a still text-only deployment. Action: Confirm whether this ID still serves text-only; if so, restore the attachment/modalities overrides (and any other true deltas).
• [medium] [possible mistake] providers/iteracompute/models/moonshotai/kimi-k3.toml:17 - Check: limit.output must be a real max completion bound. Why: output = 999_999 looks like a placeholder/sentinel, not a documented cap (lab default is 131_072; peers use concrete values). Action: Set output to the catalog/API max completion tokens for this model, or omit if inheriting lab is correct.
• [low] [possible mistake] providers/iteracompute/models/moonshotai/kimi-k3.toml:9 - Check: structured_output = false must be a verified host delta. Why: Lab metadata has structured_output = true; the same override appears on GLM-5.3 / GLM-5.3-Flash without host-specific evidence in the PR body beyond the models list. Action: Confirm IteraCompute rejects structured/JSON mode for these IDs; if not, drop the override and inherit lab.