chore(sync): update major ZenMux model families

[Alcatraz-Zhang]

Summary

• reconcile ZenMux's 82 live Qwen, Z.AI, MiniMax, xAI, and ByteDance routes against the public page allowlist
• add 23 complete canonical lab metadata files and factor every changed ZenMux route through base_model
• replace stale volcengine/* aliases with the live bytedance/* route IDs and remove 12 routes absent from the page
• update discounted USD/MTok prices, context/output limits, modalities, and host reasoning controls from the live protocol catalogs

This is the second independent split of the ZenMux refresh. It does not include the already-reviewed Claude/GPT/Gemini/Kimi/DeepSeek branch.

Sources

• https://zenmux.ai/api/frontend/model/listByFilter
• https://zenmux.ai/docs/api/openai/openai-list-models.html
• https://zenmux.ai/docs/api/anthropic/anthropic-list-models.html
• https://zenmux.ai/docs/api/vertexai/google-list-models.html
• https://zenmux.ai/docs/guide/advanced/reasoning.html
• https://huggingface.co/Qwen/Qwen3-14B
• https://huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507
• https://huggingface.co/zai-org/GLM-Image

Source snapshot: 2026-09-19. The three protocol list APIs report effective discounted prices; non-token per-count/per-second prices are intentionally omitted from the token-cost schema.

Validation

• bun validate
• bun run compare:migrations
• generator idempotence: 0 create / 0 update / 0 remove
• live/local family reconciliation: 82 / 82, no missing or extra IDs
• source-field audit: 82 models checked, 0 mismatches
• structure audit: 116 changed files, 0 missing bases, 0 full-inline provider files, 0 duplicate inherited fields, 0 incomplete metadata, 0 undocumented toggles
• git diff --check
• targeted schema/generation tests: 18 passed; the repository-wide open-weight link inventory retains its 43 pre-existing missing-link failures, none introduced by this PR

[github-actions]

Action items

• [high] [violation] providers/zenmux/models/x-ai/grok-4.2-fast.toml:2 - Check: base_model must identify the underlying lab model, and reasoning capability must match that model. Why: This route is named “Grok 4.2 Fast” but points at xai/grok-4.20-0309-non-reasoning (reasoning = false). The sibling non-reasoning route already uses that base; first-party/peers map the reasoning 4.20 surface to xai/grok-4.20-0309-reasoning with always-on reasoning_options = []. As written, the Fast route inherits non-reasoning metadata. Action: Point grok-4.2-fast at xai/grok-4.20-0309-reasoning (or the correct lab id), set reasoning / reasoning_options for this host, and keep non-reasoning only on the non-reasoning file.
• [high] [violation] providers/zenmux/models/qwen/qwen3-max.toml:4 - Check: Non-lab hosts must not force reasoning = true onto a non-reasoning lab base, and reasoning_options = [] must mean no caller control. Why: base_model = "alibaba/qwen3-max" is lab/providers/alibaba reasoning = false, but the ZenMux file renames it “Qwen3-Max-Thinking”, sets reasoning = true, and authors reasoning_options = []. Peers treat Max Thinking as a separate reasoner with a real control (e.g. OpenRouter toggle). Action: Either add/use a real thinking lab entry and author the host controls ZenMux exposes, or stop labeling/overriding this route as a reasoner on alibaba/qwen3-max.
• [high] [violation] providers/zenmux/models/z-ai/glm-5.toml:7 - Check: On multi-model relays, do not author reasoning_options = [] when the lab and same-surface peers expose caller controls. Why: Lab providers/zhipuai/glm-5 and peers such as OpenRouter/zai use toggle; ZenMux’s own provider docs document reasoning.enabled. This PR still sets [] on GLM-5 and the same pattern on glm-4.6, glm-4.7, glm-5-turbo, glm-5v-turbo, and several 4.6v/4.7-flash routes, which means “no control” rather than uncertainty. Action: Align these routes with lab/peer toggle (and wire comment) when ZenMux forwards on/off; use [] only with affirmative always-on evidence per model.
• [high] [violation] providers/zenmux/models/qwen/qwen3.5-plus.toml:12 - Check: Relay reasoning options must follow the lab/same-surface baseline, not invent GPT-style low/medium/high. Why: Lab Alibaba and DashScope-style peers use toggle + budget_tokens for Qwen3.5 Plus/Flash; this PR authors effort ["low","medium","high"] only (same shape on qwen3.5-flash). That is a different control set than the lab baseline. Action: Replace with the ZenMux-exposed Qwen surface (reasoning.enabled and/or reasoning.max_tokens / budget) matching lab/peers; do not keep L/M/H unless this host documents those exact effort values for these models.
• [high] [possible mistake] providers/zenmux/models/qwen/qwen3-asr-flash.toml:7 - Check: Provider overrides must not invent reasoning on non-reasoning lab models without host evidence. Why: New lab models/alibaba/qwen3-asr-flash and first-party Alibaba both set reasoning = false for this ASR model, but ZenMux forces reasoning = true, interleaved reasoning, and effort low|medium|high. That is a large capability change for a speech route. Action: Verify against ZenMux’s live catalog/docs; if ASR is non-reasoning, drop the reasoning override and options; if it truly reasons here, document the wire controls and why lab metadata differs.
• [medium] [possible mistake] models/zhipuai/glm-image.toml:6 - Check: Lab modalities/attachment must match the model’s described capabilities. Why: Description claims image-to-image editing and multi-subject consistency, but the lab file sets attachment = false and input = ["text"] only (image output). Downstream ZenMux inherits that and cannot represent image input. Action: Confirm GLM-Image I/O; if editing is supported, set attachment = true and include image in input modalities.
• [medium] [possible mistake] models/minimax/MiniMax-H3.toml:9 - Check: open_weights must match the model’s published weight status. Why: Description calls MiniMax H3 a “lightweight, open-weights video generation model,” but the lab file sets open_weights = false with no weights link. Action: Verify release status; set open_weights = true (and weights URL if public) or remove the open-weights claim from the description.
• [medium] [possible mistake] providers/zenmux/models/qwen/qwen3.6-plus.toml:8 - Check: When the lab exposes toggle + reasoning budget and this host’s OpenAI path supports reasoning.max_tokens, do not drop budget without evidence. Why: Lab/providers/alibaba Qwen3.6 Plus (and 3.7 Max/Plus) use toggle + budget_tokens; ZenMux keeps toggle only (3.7-flash uses []). Provider docs list reasoning.max_tokens on the OpenAI surface. Action: Confirm ZenMux catalog fields for these routes; add budget_tokens where reasoning.max_tokens is exposed, and only keep toggle-only/[] with explicit host evidence.

[Alcatraz-Zhang]

Addressed the first automated review in d0b4ea0. Added a canonical Qwen3 Max Thinking base and aligned Qwen3.5/3.6/3.7 routes to toggle + reasoning.max_tokens; replaced uncertain GLM empty controls with peer-backed toggles; added GLM-Image image input; and marked MiniMax H3 open-weight with its official Hugging Face source and license. For Grok 4.2 Fast, the underlying identity now points to the reasoning checkpoint, but the ZenMux host override remains reasoning=false because the 2026-09-19 page and both live OpenAI/Anthropic lists explicitly report false for that exact route. For Qwen3 ASR, both ZenMux page and OpenAI list explicitly report reasoning=true, but no reasoning parameter or side channel is advertised, so it now uses [] and no interleaved field instead of the earlier invented L/M/H controls. Revalidated 117 files, exact 82/82 live set, 82 source rows with zero mismatches, all base/override/header checks, bun validate, compare:migrations, and diff check.

[github-actions]

Action items

• [high] [possible mistake] providers/zenmux/models/bytedance/doubao-seed-1.8.toml:8 - Check: Relay reasoning_options must match this host’s real controls and the lab/same-surface peer baseline for the model, not an invented narrower set. Why: First-party Volcengine entries for Seed 1.8 / 2.0 lite / mini / pro use effort none|minimal|low|medium|high|xhigh|max (and interleaved.reasoning_content). This PR sets only low|medium|high on doubao-seed-1.8, doubao-seed-2.0-lite, doubao-seed-2.0-mini, doubao-seed-2.0-pro, and doubao-seed-code, while sibling ZenMux routes (doubao-seed-2.0-code, doubao-seed-2.1-*, doubao-seed-evolving) use the full seven-value set—internally inconsistent for one host. Action: Verify ZenMux’s live OpenAI list / reasoning docs for each of these IDs; if the host exposes the full Seed enum (or passthrough), replace low|medium|high with the lab set and add interleaved where the side channel is present. If the host truly only accepts L/M/H, document that host-only wire constraint in the leading comment and keep the narrow list consistently.
• [medium] [possible mistake] providers/zenmux/models/qwen/qwen3.8-max.toml:34 - Check: After base_model, modality overrides must be real host deltas, not accidental drops of lab capabilities. Why: Lab models/alibaba/qwen3.8-max.toml and qwen3.8-max-0902.toml include pdf input; both ZenMux files override to ["text", "image", "video"] only, dropping PDF without a cited host limitation. Action: Confirm whether ZenMux’s Qwen3.8 Max routes accept PDF; restore pdf if supported, or keep the override with a leading comment stating the host omits PDF.
• [medium] [possible mistake] providers/zenmux/models/bytedance/doubao-seed-1.8.toml:35 - Check: Provider modality overrides vs lab base. Why: Lab bytedance-seed/seed-1-8 is input = ["text", "image"]; ZenMux adds video. Same pattern: doubao-seed-2.0-lite / mini add audio beyond lab text|image|video. Action: Confirm these are live ZenMux modality catalogs; if not, drop the extras so inheritance matches the lab files.
• [low] [possible mistake] models/alibaba/qwen3-14b.toml:5 - Check: New lab metadata should use accurate release/last_updated dates when known. Why: release_date / last_updated / knowledge are month-only (2025-04). Month form is schema-valid, but peers and HF-backed entries usually carry full days; imprecise dates weaken catalog quality. Action: Prefer the public release day (e.g. from the linked Hugging Face card) if known; otherwise leave as-is.

[Alcatraz-Zhang]

Addressed the second automated review in 8190241. The five older ByteDance reasoning routes now use the full peer-backed effort enum none|minimal|low|medium|high|xhigh|max plus reasoning_content, matching newer ZenMux Seed routes. Kept the host modality deltas because both live OpenAI and Anthropic lists explicitly report video for Seed 1.8 and video+audio for Seed 2.0 Lite/Mini; leading comments now record that evidence. Likewise, both live lists explicitly omit PDF for Qwen3.8 Max/0902, so those host overrides remain and are documented. Qwen3-14B dates are now the full page-backed 2025-04-28. Revalidated generator idempotence, 117-file structure audit, exact 82/82 live set, 82 source rows with zero mismatches, bun validate, compare:migrations, and diff check.

[github-actions]

Action items

• [high] [possible mistake] providers/zenmux/models/qwen/qwen3.8-flash.toml:30 - Check: Provider limit.output must be a real host max-completion bound, not the context window. Why: This route sets output = 1_000_000 while lab/first-party Qwen3.8 Flash resolve around 131_072 output; a 1M completion cap is almost certainly the context figure copied into output and will mislead clients on max generation size. Action: Verify ZenMux’s live max output for qwen3.8-flash and set limit.output to that value (or drop the override and inherit the lab default).
• [medium] [possible mistake] providers/zenmux/models/x-ai/grok-build-0.1.toml:3 - Check: reasoning overrides on relays must match this host’s surface, with evidence when contradicting lab/first-party. Why: Lab + xAI first-party mark Grok Build 0.1 as reasoning = true (always-on []), but ZenMux sets reasoning = false with no wire/catalog note (unlike the explicit supports_reasoning=0 rationale on grok-4.2-fast). Action: Confirm ZenMux’s list APIs for this id; either keep reasoning = true + reasoning_options = [], or document the host-disabling evidence in the leading comment.
• [medium] [possible mistake] providers/zenmux/models/qwen/qwen3-asr-flash.toml:8 - Check: Provider reasoning should not invent a reasoner when lab/first-party ASR is non-reasoning unless the host truly exposes a reasoning surface. Why: models/alibaba/qwen3-asr-flash and providers/alibaba/.../qwen3-asr-flash both use reasoning = false, yet ZenMux forces reasoning = true with reasoning_options = [] on a speech-transcription route. Empty options mean “always on, no control,” which is an odd claim for ASR. Action: Re-check the live OpenAI/Anthropic entries; if there is no reasoning side channel or control, set reasoning = false and drop reasoning_options.
• [medium] [possible mistake] providers/zenmux/models/qwen/qwen3.6-plus.toml:28 - Check: Modality overrides must be real host deltas, not broader than lab/first-party without evidence. Why: Final file advertises input = ["text", "image", "pdf", "video"], but lab and Alibaba first-party Qwen3.6 Plus are text/image/video only (no PDF). Action: Confirm ZenMux’s live modality list for this route; remove pdf if not actually accepted, or cite the host field that proves PDF input.
• [medium] [possible mistake] models/alibaba/qwen3-max-thinking.toml:1 - Check: New models/ entries must be real lab identities, not host-only thinking modes of an existing model. Why: Canonical Alibaba metadata is alibaba/qwen3-max with reasoning = false; no first-party qwen3-max-thinking lab/provider entry exists. ZenMux’s route id remains qwen/qwen3-max while pointing base_model at a new synthetic lab id. Action: Prefer base_model = "alibaba/qwen3-max" plus provider overrides (reasoning, reasoning_options, display name), or prove Alibaba ships a distinct Max-Thinking model id before keeping the new lab file.
• [low] [possible mistake] providers/zenmux/models/z-ai/glm-image.toml:4 - Check: Host modality/attachment overrides should not silently drop documented lab capabilities without a route-specific reason. Why: After correcting the lab entry to attachment = true and input = ["text", "image"] (image-to-image editing), the ZenMux file overrides back to attachment = false and text-only input. Action: Verify ZenMux’s GLM-Image route; restore image input/attachment if the API accepts image edits, or comment why this host is text-to-image only.

[Alcatraz-Zhang]

Addressed the third automated review in 8fe7c50. Qwen3 Max now uses the existing alibaba/qwen3-max base with ZenMux-only name/reasoning/control overrides, and the synthetic lab identity is removed. Qwen3 ASR is now non-reasoning because its specialized transcription surface exposes neither a reasoning request control nor response channel despite the upstream list flag. Kept Qwen3.8 Flash output=1,000,000 because the 2026-09-19 ZenMux page explicitly reports max_completion_tokens=1000000; kept Qwen3.6 Plus PDF because both live OpenAI/Anthropic lists explicitly include file input; kept GLM-Image text-only because the live Google list exposes only text on this hosted route; and documented Grok Build reasoning=false from the page plus both live protocol lists. Revalidated generator idempotence, 116-file structure audit, exact 82/82 set, 82 source rows with zero mismatches (including the documented ASR semantic correction), bun validate, migration comparison, and diff check.

[github-actions]

Action items

• [medium] [possible mistake] providers/zenmux/models/qwen/qwen3-max.toml:5 - Check: Provider base_model identity must preserve real host product facts (name/dates/description) when the route is not the plain lab non-reasoning Max. Why: Final state points at alibaba/qwen3-max (reasoning = false, release_date/last_updated = 2025-09-23, Max-oriented description) while the ZenMux route is still named Qwen3-Max-Thinking with reasoning = true. Patch 2’s dedicated models/alibaba/qwen3-max-thinking.toml (dates 2026-01-23) was deleted in patch 4, so this route silently inherits Max’s earlier launch metadata and non-thinking description despite the Thinking display name and reasoning overrides. Action: Either restore a complete thinking lab entry and base_model it, or keep alibaba/qwen3-max and explicitly override description, release_date, and last_updated (and any other Thinking-only facts) so the resolved catalog does not publish Max’s non-thinking identity under a Thinking name.
• [medium] [possible mistake] providers/zenmux/models/z-ai/glm-5.2.toml:9 - Check: Relay reasoning side-channel should match the lab / same-host peer surface when this host exposes reasoning_content. Why: First-party providers/zhipuai/models/glm-5.2.toml and OpenRouter’s glm-5.2 both set [interleaved] field = "reasoning_content", and this PR’s sibling ZenMux routes (glm-5.3, glm-5.3-flash, glm-5.3-flashx, glm-5, glm-5.1) do the same, but glm-5.2 only authors effort = high|max with no interleaved field. Action: Add [interleaved] field = "reasoning_content" for ZenMux glm-5.2 unless the live protocol catalogs show this one route has no reasoning side channel (in that case, document that host exception in the leading comment).

[Alcatraz-Zhang]

Addressed the fourth automated review in af297ca and merged the latest upstream/dev (dff014f) first. Qwen3 Max continues to use the existing alibaba/qwen3-max base, but now explicitly overrides the ZenMux Thinking-route description, release_date, and last_updated so it no longer inherits the plain Max identity. GLM-5.2 now exposes interleaved.reasoning_content, matching first-party and same-host peers. Revalidated against the new dev base: generator idempotence, 116-file structure audit, exact 82/82 live set, 82 source rows with zero mismatches, bun validate, compare:migrations, and diff check.

[github-actions]

Action items

• [high] [violation] providers/zenmux/models/qwen/qwen3-max.toml:7 - Check: Provider model metadata must be internally consistent and must not contradict the authored reasoning surface. Why: Final file names the route Qwen3-Max-Thinking, sets reasoning = true, and authors toggle + budget_tokens, but the new description still says the model “does not include a dedicated ‘thinking’ mode.” That contradicts the same file’s capability fields and the lab base (alibaba/qwen3-max is reasoning = false). Consumers will get mixed signals about whether this host exposes thinking controls. Action: Resolve the identity: either (a) keep the thinking surface and rewrite description (and dates/name) to match a real reasoning-capable route, preferably via a distinct complete lab entry if this is not plain qwen3-max, or (b) drop reasoning/reasoning_options and the “Thinking” branding if ZenMux truly serves non-thinking qwen3-max.

[Alcatraz-Zhang]

Addressed the fifth automated review in 22588eb. The live ZenMux entry itself is contradictory: it is named Qwen3-Max-Thinking and reports reasoning=true, but its source description ends by denying a thinking mode. The provider description now preserves the route’s verified use cases while explicitly identifying it as ZenMux’s reasoning-capable Qwen3 Max route, so the description, name, reasoning flag, and toggle/budget controls are internally consistent. Revalidated generator idempotence (0 changes), 116-file structure audit, exact 82/82 live set, 82 source rows with zero mismatches, bun validate, compare:migrations, and diff check.

[github-actions]

Action items

• [high] [violation] providers/zenmux/models/qwen/qwen3-max.toml:6 - Check: Non-lab hosts must base_model the underlying lab identity; if that lab entry is missing, add a complete models/<lab>/<id>.toml rather than inheriting a different model. Why: This route is authored as Qwen3-Max-Thinking with reasoning = true plus toggle/budget controls, but base_model = "alibaba/qwen3-max" points at the non-reasoning Qwen3 Max lab entry (reasoning = false). A distinct Max Thinking product is already nameable elsewhere (e.g. OpenRouter qwen3-max-thinking), and this PR previously added then deleted models/alibaba/qwen3-max-thinking.toml. Merging thinking-route metadata onto the non-thinking lab identity mislabels the catalog and forces capability overrides that belong on a real lab base. Action: Restore a complete models/alibaba/qwen3-max-thinking.toml (or equivalent lab id matching the live product) and point this ZenMux file at it as an override-only host entry; drop the forced reasoning = true / date restatements that only exist to paper over the wrong base.
• [medium] [possible mistake] providers/zenmux/models/qwen/qwen3-asr-flash.toml:11 - Check: Provider limit overrides must reflect the host route and stay consistent with the lab ASR identity. Why: Lab/first-party Qwen3 ASR Flash uses context = 53_248 / output = 4_096. This host entry still overrides to context = 1_000_000 / output = 65_536 after the reasoning cleanup—limits that look like generic LLM defaults, not speech-transcription bounds. Action: Confirm against ZenMux’s live list for this id; if the page does not actually advertise ~1M/65k for ASR, restore lab limits (or only author the verified host deltas).
• [medium] [possible mistake] providers/zenmux/models/minimax/minimax-m2-her.toml:4 - Check: reasoning overrides on relays need the same host-evidence bar used for other ZenMux reasoning deltas. Why: Lab minimax/MiniMax-M2-Her is reasoning = true, but this route sets reasoning = false with no leading comment. Peer MiniMax routes in the same PR keep inherited reasoning ([]). Other ZenMux non-reasoning overrides (e.g. Grok Build / Grok 4.2 Fast) document live supports_reasoning / capabilities flags. Action: Either document host evidence that this id is non-reasoning on ZenMux, or remove the override and author reasoning_options if the host still exposes reasoning.
• [low] [possible mistake] providers/zenmux/models/x-ai/grok-imagine-image-2.0.toml:6 - Check: Limit overrides for image routes should match the host catalog, not an unexplained jump from lab defaults. Why: Lab xai/grok-imagine-image-2.0 has context = 8_000; this file overrides to 66_000 while only commenting that pricing is non-token. Action: Cite the ZenMux list field that justifies 66k, or drop the override and inherit the lab limit.

[Alcatraz-Zhang]

Addressed the sixth automated review in d7f3703. Qwen3 Max now uses the real Alibaba API snapshot identity alibaba/qwen3-max-2026-01-23, backed by complete lab metadata and Alibaba Model Studio/Qwen release sources; the ZenMux route is override-only for its hosted display name, controls, pricing, structured output, and limits. This avoids restoring the non-API synthetic qwen3-max-thinking id while satisfying the requested equivalent lab identity. Verified the other three host deltas against the live ZenMux sources and documented them in leading comments: Qwen3 ASR Flash is explicitly 1,000,000 context / 65,536 completion on the page (with OpenAI independently reporting 1,000,000 context); MiniMax M2-her reports supports_reasoning=0 and OpenAI reasoning=false; Grok Imagine Image 2.0 reports 66,000 context on both page and OpenAI list. Revalidated generator idempotence (0 changes), 117-file structure audit, exact 82/82 live set, 82 source rows with zero mismatches, bun validate, compare:migrations, and diff check. Official Qwen references: https://docs.modelstudio.console.alibabacloud.com/en/model-studio/model-qwen3-max and https://www.alibabacloud.com/blog/pushing-qwen3-max-thinking-beyond-its-limits_602834

[github-actions]

No actionable findings.

[rekram1-node]

this PR is massive can u plz split it up

[Alcatraz-Zhang]

Agreed — I am splitting this into three independent, non-stacked PRs from the latest dev so each family group is small and reviewable. The first replacement is #7481 (Qwen: 41 files). The remaining groups will be Z.AI+xAI (35 files) and ByteDance+MiniMax (41 files), opened sequentially after each automated review clears. This original PR is now Draft and will remain only as the reviewed source snapshot until all replacements are open.

[Alcatraz-Zhang]

Superseded by three independent, non-stacked PRs from the same current dev base, as requested in human review: #7481 (Qwen, 41 files), #7484 (Z.AI + xAI, 35 files), and #7485 (ByteDance + MiniMax, 41 files). I verified the GitHub file lists: their union is exactly the original 117-file diff, with zero overlaps, omissions, or extras. Closing this large source PR; its branch and review history remain available for traceability.