feat(layerx1): add LayerX1 provider with 32 models

[ATOM00blue]

Adds LayerX1 (layerx1.com · docs) as a provider.

LayerX1 is an OpenAI-compatible multi-model gateway. The same key also serves the
Anthropic Messages and OpenAI Responses wires, so @ai-sdk/openai-compatible +
api is the right entry.

Contents

• provider.toml — LAYERX1_API_KEY, api = https://api.layerx1.com/v1
• logo.svg — currentColor only, no hardcoded colors, no fixed width/height, square viewBox
• 32 model TOMLs under models/

How the model files were authored

LayerX1 hosts models it did not create, so every file uses base_model against
the existing lab entry and is override-only. No lab metadata (name,
description, dates, open_weights, matching limit/modalities) is restated.

Provider-authored fields are cost and reasoning_options. Beyond those, the only
values present are genuine deltas where LayerX1's published capability differs from
the lab entry — a smaller served context, or a text-only lane for a model that is
multimodal upstream.

Costs come from the live catalog at https://api.layerx1.com/v1/models
(pricing_usd_per_mtok), with cached_input mapped to cache_read.

reasoning_options

These are uniform across the reasoning models, which is intentional rather than
copy-paste. The gateway normalizes every backend onto a single neutral effort
ladder — none | minimal | low | medium | high | max — and exposes reasoning
visibility separately through include_reasoning, independent of effort.
Anthropic-style token budgets are accepted and translated onto the same control.
So the surface a caller sees does not vary per backend:

reasoning_options = [
  { type = "toggle" },
  { type = "effort", values = ["none", "minimal", "low", "medium", "high", "max"] },
  { type = "budget_tokens" },
]

xhigh is deliberately absent — the schema permits it but LayerX1 does not
implement it, and the ladder is a documented API surface rather than an inference.

Deliberately not included

• 7 embedding models (BGE family, embeddinggemma-300m, qwen3-embed-0.6b,
  plamo-embed-1b). These need lab entries that don't exist yet — there is no
  models/baai/ at all. Happy to open a follow-up PR authoring complete lab
  metadata for those rather than bundle it here.
• 2 models whose upstream identity I could not establish with confidence.
  Left out rather than guessed at a base_model.

Validation

bun validate passes, and layerx1 resolves with all 32 models.

Offer

LayerX1's /v1/models is a rich catalog endpoint — it returns pricing, context
window, capabilities, and per-model reasoning support, and it is authoritative
about removals. That fits the sync-module criteria in sync.md, so if you'd
prefer this entry be sync-maintained rather than hand-authored, I'm glad to add
the module in a follow-up.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[github-actions]

Action items

• [high] [violation] providers/layerx1/models/lx1-deepseek-v3.2.toml:4 - Check: toggle must not be paired with effort that already includes none. Why: Every reasoning model uses { type = "toggle" } together with effort values that include "none" (e.g. deepseek, glm, gpt-oss, grok, kimi, longcat, minimax, nemotron, opus, qwen, sonnet). AGENTS.md / the reasoning-options audit treat that as invalid: off is either effort=none or a separate toggle, not both. The PR’s include_reasoning “visibility” control is not a valid reason to keep both shapes. Action: Pick one wire model and apply it consistently: either drop toggle and keep effort with none, or drop none from effort and keep toggle only if it is a true on/off control (not response visibility). Update all ~24 reasoning TOMLs the same way.
• [high] [violation] providers/layerx1/models/lx1-deepseek-v3.2.toml:1 - Check: Every toggle needs a leading top-of-file wire comment on the model file. Why: Reasoning model files start with base_model and declare toggle with no per-file wire path. A note only in provider.toml does not satisfy the requirement (and sync-safe comments must sit above the first key in the file that owns the toggle). Action: If toggle remains, add a leading comment on each toggled model file with the exact request field (e.g. include_reasoning / effort path). If toggle is removed per the item above, this is moot.
• [medium] [possible mistake] providers/layerx1/models/lx1-deepseek-v3.2.toml:6 - Check: budget_tokens only when this host exposes a real reasoning-token budget field. Why: The same budget_tokens entry is stamped on every reasoner, including surfaces that are effort-only upstream (DeepSeek V4, Claude 4.8+/5 adaptive effort, gpt-oss, etc.). Translating Anthropic budgets into an effort ladder is not by itself a budget control; advertising budget_tokens tells clients a integer reasoning budget is first-class. Action: Keep budget_tokens only if LayerX1’s API actually accepts a reasoning budget parameter with meaningful effect; otherwise remove it. Cite the exact wire field in a leading comment when kept.
• [medium] [possible mistake] providers/layerx1/models/lx1-kimi-k2.5.toml:3 - Check: Provider reasoning overrides must match what this host actually serves. Why: reasoning = false contradicts lab models/moonshotai/kimi-k2.5.toml and first-party providers/moonshotai/models/kimi-k2.5.toml (reasoning = true with a toggle). Forcing non-reasoning drops required reasoning_options and mislabels a thinking model if the gateway still runs it as one. Action: Verify LayerX1’s live behavior/docs for lx1-kimi-k2.5; restore reasoning = true + correct reasoning_options, or document evidence that this lane is non-thinking only.
• [medium] [possible mistake] providers/layerx1/models/lx1-qwen-turbo.toml:3 - Check: Provider reasoning overrides must match what this host actually serves. Why: reasoning = false contradicts lab models/alibaba/qwen-turbo.toml and first-party Alibaba (reasoning = true, toggle + budget_tokens). Action: Confirm LayerX1 disables thinking for this ID; if not, drop the override and author reasoning_options for the host surface.
• [medium] [possible mistake] providers/layerx1/models/lx1-opus-5.toml:3 - Check: Text-only / no-attachment overrides should be complete and consistent. Why: lx1-opus-5 and lx1-sonnet-5 set attachment = false but still inherit lab [modalities].input of text/image/pdf. Sibling lx1-opus-4.8 correctly pairs attachment = false with input = ["text"]. Callers will still see multimodal input support on Opus 5 / Sonnet 5. Action: If these lanes are text-only like Opus 4.8, also override [modalities].input = ["text"] (and match any other real capability deltas); if vision/PDF still work, drop the incomplete attachment = false override or justify it.

[ATOM00blue]

Thanks — the reasoning-options criticism was correct on all counts. Fixed in cff2408.

1 & 2, toggle paired with effort containing none (high). Agreed, and I've
dropped toggle rather than keep both shapes. LayerX1 has no separate boolean enable:
reasoning is requested through a single reasoning effort parameter whose none value
is the off switch, so a toggle duplicated effort = none. You were also right to
reject the include_reasoning defence — that controls whether reasoning text is
returned, not whether the model reasons, so it does not justify a second control. All
reasoning files now carry exactly one shape, which makes item 2 moot. I removed that
claim from provider.toml too, since it was no longer load-bearing.

3, budget_tokens (medium). Removed. Checked /v1/models: no budget or thinking
parameter is advertised for any model — the only reasoning-related entry in
supported_parameters is reasoning. Anthropic-style budgets are accepted on the
inbound Anthropic wire but collapse onto the effort ladder, which as you say is not a
budget control. Declaring it would have overstated the surface.

4 & 5, reasoning = false on lx1-kimi-k2.5 and lx1-qwen-turbo (medium). These
are intentional and I've cited the evidence inline. /v1/models omits reasoning from
both capabilities.reasoning and supported_parameters for these two ids, while
sibling reasoning lanes such as lx1-deepseek-v3.2 list it in both. So the host's own
declaration is that these lanes are served non-thinking, independent of upstream
capability. Happy to drop the ids entirely if you would rather not carry a
non-thinking lane for a thinking model.

6, incomplete text-only overrides (medium). Real bug, thank you. Rather than patch
the two files I fixed the rule that produced them: attachment is now true when the
lane accepts any file, so it can no longer contradict a vision-capable model, and
modalities.input is filtered by capability — image/video gate on vision, pdf
gates on documents. Your example generalised further than the two models you flagged:
the same fault had left video on the Kimi lanes while they declared
attachment = false.

Net effect: lx1-opus-5 and lx1-sonnet-5 are now ["text", "image"] with
attachment inherited true; lx1-opus-4.8 is unchanged at ["text"]. I also added an
assertion over all 32 models that attachment, every modalities.input entry, and
reasoning agree with the capability flags /v1/models reports — it now reports zero
inconsistencies, and bun validate passes.

[github-actions]

Action items

• [high] [possible mistake] providers/layerx1/models/lx1-qwen3-235b.toml:1 - Check: Provider base_model must name the actual lab model, and host limits must match that identity. Why: base_model = "alibaba/qwen3-235b-a22b" is the original hybrid reasoner (reasoning = true, lab limit.context = 131_072), but this file sets context = 262_144 — the lab context of alibaba/qwen3-235b-a22b-instruct-2507 (reasoning = false). That mix risks the wrong lab identity and wrong capability inheritance. Action: Confirm against LayerX1 /v1/models which checkpoint lx1-qwen3-235b is. Point base_model at that lab entry (add it if missing), and keep only real host deltas for limit / reasoning / reasoning_options.
• [medium] [possible mistake] providers/layerx1/models/lx1-qwen3-next-80b.toml:8 - Check: Host limit.context overrides should reflect the served window, not exceed the lab entry without evidence. Why: Lab alibaba/qwen3-next-80b-a3b-instruct is 131_072, but this file raises context to 262_144. Action: Verify 262_144 from LayerX1’s catalog/docs for this id; if the host still serves 131k, drop the override (or correct the value).
• [medium] [possible mistake] providers/layerx1/models/lx1-minimax-m2.5.toml:12 - Check: Host output limits should match the provider’s published max generation size. Why: Lab minimax/MiniMax-M2.5 has output = 131_072, but this file sets output = 8_192 (and trims context to 196_608). That is a large capability cut and will mislead clients if the API allows more. Action: Confirm both limits from LayerX1 /v1/models (or docs) for lx1-minimax-m2.5 and correct them if the catalog differs.
• [low] [possible mistake] providers/layerx1/models/lx1-deepseek-v3.2.toml:14 - Check: Host limits should not silently exceed the lab baseline without a catalog source. Why: Lab deepseek/deepseek-v3.2 is context = 128_000, but this file sets 131_072. Same pattern on lx1-grok-4.3 (output = 32_000 vs lab 30_000). Action: Verify these values on LayerX1’s model catalog and align the TOMLs (or cite the exact catalog fields in a leading comment if the higher host limits are intentional).