feat(melious): add Melious provider with the GLM, DeepSeek and Kimi families

[jope35]

Adds Melious (Melious AI GmbH), a European OpenAI-compatible
relay of open-weight models, with an endpoint at https://api.melious.ai/v1.

Melious is a multi-model relay, not a lab — it did not build any model it serves —
so every entry uses base_model and is override-only.

Scope: first of two PRs

This PR is deliberately narrow: the provider plus the GLM, DeepSeek and Kimi
families (15 models), all of which already have lab metadata under models/.

It touches providers/melious/ only — no models/, no labs/, no
packages/web/src/render.tsx. A second PR adds the rest of the catalog, including the
11 models that need new lab entries (BAAI, intfloat, sentence-transformers, NousResearch,
H Company and two Qwen ids). Splitting it this way keeps the shared-metadata discussion
out of the provider's own review.

Data provenance

Primary source is the authenticated catalog, GET https://api.melious.ai/v1/models?include_meta=true,
cross-checked against the per-model hub pages (https://melious.ai/hub/models/<id>),
both accessed 2026-09-14.

Prices are the balanced route: a bare chat model ID resolves to balanced
(routing docs), so that is what a caller of
model="glm-5.3" is billed. Only bare IDs are published — no :suffix variants.

Converted at 1.1592 USD/EUR, the ECB euro reference rate for 2026-09-11.
One rate for the whole PR; every file's leading comment records its EUR figures, the rate
and the date, so the arithmetic is checkable from the diff without leaving the page.

model	base_model	context	EUR in/out/cache	USD in/out/cache
deepseek-r1-0528	deepseek/deepseek-r1	164,000	0.6 / 2.4 / 0.16	0.69552 / 2.78208 / 0.185472
deepseek-v3.2	deepseek/deepseek-v3.2	164,000	0.3 / 0.5 / 0.08	0.34776 / 0.5796 / 0.092736
deepseek-v4-flash-0731	deepseek/deepseek-v4-flash-0731	1,000,000	0.1 / 0.25 / 0.02	0.11592 / 0.2898 / 0.023184
deepseek-v4-pro	deepseek/deepseek-v4-pro	1,000,000	1.6 / 3.2 / 0.4	1.85472 / 3.70944 / 0.46368
deepseek-v4-pro-0813	deepseek/deepseek-v4-pro-0813	1,000,000	1.0 / 3.0 / 0.1	1.1592 / 3.4776 / 0.11592
deepseek-v4.1-flash	deepseek/deepseek-v4.1-flash	1,000,000	0.2 / 1.0 / 0.01	0.23184 / 1.1592 / 0.011592
glm-5	zhipuai/glm-5	203,000	0.95 / 2.9 / 0.23	1.10124 / 3.36168 / 0.266616
glm-5.1	zhipuai/glm-5.1	203,000	1.3 / 4.0 / 0.32	1.50696 / 4.6368 / 0.370944
glm-5.2	zhipuai/glm-5.2	1,000,000	1.0 / 4.0 / 0.25	1.1592 / 4.6368 / 0.2898
glm-5.3	zhipuai/glm-5.3	1,000,000	1.0 / 3.0 / 0.2	1.1592 / 3.4776 / 0.23184
glm-5.3-flash	zhipuai/glm-5.3-flash	1,000,000	0.1 / 0.4 / 0.02	0.11592 / 0.46368 / 0.023184
kimi-k2.5	moonshotai/kimi-k2.5	262,000	0.5 / 2.55 / 0.12	0.5796 / 2.95596 / 0.139104
kimi-k2.6	moonshotai/kimi-k2.6	256,000	0.7 / 3.5 / 0.23	0.81144 / 4.0572 / 0.266616
kimi-k2.7-code	moonshotai/kimi-k2.7-code	262,144	0.7 / 3.0 / 0.19	0.81144 / 3.4776 / 0.220248
kimi-k3	moonshotai/kimi-k3	1,000,000	2.75 / 13.75 / 0.68	3.1878 / 15.939 / 0.788256
Worth noting for reviewers: the hub's rendered pages disagree with the catalog endpoint on
six of these — kimi-k2.7-code output is EUR 3.00 in the catalog but EUR 3.50 on the hub,
kimi-k2.5 context is 262 000 not 256 000, and four cache rates are rounded in the hub UI.
The catalog figures are used throughout.

Reasoning options

reasoning_effort is the only reasoning control this host exposes. It forwards no
lab-native thinking field — no thinking.type, no enable_thinking, no reasoning budget —
so no entry uses toggle or budget_tokens.

The host accepts any effort string without a 4xx, including minimal, xhigh and
max, so acceptance alone proves nothing. The authored set is therefore the documented
ladder plus one measured value:

reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high"] }]

• none is a real off state — measured, not assumed. reasoning_effort="none" returns
  usage.reasoning_tokens = 0 and no reasoning_content, reproduced on glm-5.3,
  deepseek-v4-pro and kimi-k2.7-code (2026-09-14, against
  api.melious.ai/v1/chat/completions). Per AGENTS.md, because none is reachable this
  is an effort list containing none, not a toggle.
• low/medium/high are the ladder Melious documents in the
  chat-completions reference.
• minimal/xhigh/max are accepted but undocumented, and produced no behaviour
  distinguishable from high, so they are not authored — no full-enum dump.

On divergence from the lab enums: several of these models expose low|high|max or a
toggle on their first-party host, and those controls are simply not reachable through this
API. Per AGENTS.md → Reasoning options, a relay's options describe its own request
surface. There is merged precedent for exactly this on same-surface relays, e.g.
providers/cortecs/models/deepseek-v4-pro.toml,
providers/cortecs/models/deepseek-v4-flash-0731.toml,
providers/digitalocean/models/deepseek-v4-flash-0731.toml and
providers/nebius/models/zai-org/GLM-5.2.toml.

[interleaved] field = "reasoning_content" is set from observed live responses, not assumed.

Overrides

Inherited from the lab entries unless Melious genuinely differs. Real deltas only:

• limit.context where the served window differs: glm-5/glm-5.1 (203 000),
  deepseek-r1-0528/deepseek-v3.2 (164 000), kimi-k2.5 (262 000), kimi-k2.6 (256 000),
  kimi-k3 (1 000 000 vs the lab's 1 048 576).
• limit.output only where the lab's output exceeds Melious's served context
  (kimi-k2.5, kimi-k2.6). max_output_tokens is null for all 15, so output is otherwise
  inherited rather than invented.
• modalities.input where this host routes less than the lab supports: glm-5.3-flash
  (no PDF) and kimi-k2.5 (no video).

Names, descriptions, dates, tool_call, open_weights and matching limits/modalities are
not restated.

Prior art

#5047 and
#5221 both propose this provider and
are both currently blocked. Their published data also contradicts each other and the live
catalog, so nothing here is ported from either — every value was re-sourced on 2026-09-14.
Two concrete corrections that came out of that: #5221 states an explicit
reasoning_effort: "none" still returns reasoning and therefore omits it — measurement above
shows it does reach an off state; and mistral/voxtral-small-2507, used as a base_model
target in #5047, does not exist (models/mistral/voxtral-small-24b-2507.toml does).

Test plan

• bun validate — exits 0
• bun run test in packages/sdk — 23 pass, 0 fail
• cd packages/web && bun run build — dist/_api.json contains melious with 15 models; dist/logos/melious.svg emitted; the provider page renders
• Every entry uses base_model; no restated-identical fields; no id key; no context_over_200k
• All comments sit above the first key, so sync re-serialization preserves them
• No limit.output > limit.context; no cache_read >= input; every resolved model has limit.context + limit.output
• Cost arithmetic reproduces from the EUR figures and rate stated in each file
• logo.svg — square viewBox, currentColor, no fixed width/height, no hardcoded colors

Happy to add a sync module if maintainers want to add a MELIOUS_API_KEY secret to
sync-models.yml — the catalog endpoint is rich enough to drive one.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] providers/melious/models/deepseek-v4-pro.toml:22 - Check: Relay reasoning_options must follow the lab/same-surface peer baseline for each model, not a universal GPT low/medium/high ladder (AGENTS.md → Reasoning options; audit-reasoning-options). Why: Melious is a multi-model openai-compatible relay, but every one of the 15 models is authored as effort = ["none", "low", "medium", "high"]. That is the GPT-gateway default and conflicts with established baselines the PR itself cites: DeepSeek V4 lab/peers use high/max (or low/high/max), not L/M/H; GLM-5.3 peers use low/high/max; GLM-5.2 is high/max; GLM-5/5.1 and several Kimi/DeepSeek peers are toggle-only or []. The files admit the host accepts any effort string without 4xx (so acceptance is not evidence), that max/xhigh/minimal showed no distinct behavior, and that live checks covered only glm-5.3, deepseek-v4-pro, and kimi-k2.7-code—mainly proving none turns reasoning off. Publishing a single invented L/M/H set misrepresents caller controls for non-GPT families. Action: Author per-model (or per-family) reasoning_options from this host’s real effective controls, starting from lab + peer sets: e.g. DeepSeek V4 → none + high (add max only if it is distinct here); GLM-5.3/5.3-flash → none + low/high/max if those grade here; toggle-only peers → none plus only levels that actually change behavior (not a full L/M/H dump); peer-[] models → only levels with host evidence. Drop medium (and any other value) unless Melious docs or measurements show a real graded effect on that model. Keep the effort+none shape (no toggle) only where off is truly reasoning_effort="none".

[jope35]

Fixed in 6180a3c — the finding was right, and the uniform ladder is gone.

You were correct on the substance: acceptance was not evidence, the live checks covered
only three models, and publishing one L/M/H set across DeepSeek, GLM and Kimi asserted a
control that was never demonstrated. Rather than re-derive from the lab/peer enums, I
measured every model on this host, because those enums are not reachable through this API
and AGENTS.md → Reasoning options asks for the host's own surface.

Method. Each of the 15 models, same prompt, temperature: 0, max_tokens: 3000 (large
enough that nothing truncated mid-trace), comparing emitted reasoning_content across
none/low/medium/high, with repeat samples wherever the result was load-bearing.
Measured 2026-09-14 against api.melious.ai/v1/chat/completions. reasoning_content
length was used rather than usage.reasoning_tokens, because the latter is unreliable here
— glm-5.2 reports reasoning_tokens: 0 while returning 2 546 characters of reasoning.

Result — two behaviours, neither of them a graded ladder.

model	none	low	medium	high	authored
glm-5	2444 / 2239 / 686 / 3907	4130	4130	2155	[]
deepseek-r1-0528	2786 / 3010	3379	4383	2918	[]
kimi-k2.5	1177 / 1714 / 1100	959	983	1265	[]
glm-5.1	0	2573	2670	2591	["none","high"]
glm-5.2	0	2843	2546	1073	["none","high"]
glm-5.3	0	88	621	174	["none","high"]
glm-5.3-flash	0	79	886	174	["none","high"]
deepseek-v3.2	0 / 0	838	836	965 / 950	["none","high"]
deepseek-v4-flash-0731	0	831	641	574	["none","high"]
deepseek-v4-pro	0	682	619	614	["none","high"]
deepseek-v4-pro-0813	0	805	678	590	["none","high"]
deepseek-v4.1-flash	0	298	277	319	["none","high"]
kimi-k2.6	0	889	936	808	["none","high"]
kimi-k2.7-code	0	297	387	295	["none","high"]
kimi-k3	0	418	301	499	["none","high"]

(reasoning_content characters; / separates repeat samples.)

• No model grades. low/medium/high is non-monotonic everywhere and mostly within
  run-to-run noise. On glm-5, low and medium returned byte-identical output at
  temperature 0 — the two values collapse to the same backend setting. So no graded levels
  are published anywhere, exactly as you asked.
• 12 models → effort ["none", "high"], the shape you proposed for DeepSeek V4.
  "none" is a verified off state; one on-level is published because only one is
  distinguishable.
• 3 models → []. On glm-5, deepseek-r1-0528 and kimi-k2.5, "none" does not
  suppress reasoning across repeat samples and nothing grades, so there is no caller
  control at all. This is an affirmative measurement rather than an untested default.

Two things worth recording for future syncs, both now in the file comments:

• My earlier none claim was generalised from three models and is simply false on three
  others — it is per-model, which is why this is now measured per model.
• The catalog's own reasoning_type does not predict behaviour: glm-5.3-flash is typed
  reasoning (always-on) yet honours "none", while glm-5 is typed hybrid yet ignores
  it. Neither vendor docs, catalog metadata, nor peer baselines were reliable here.

bun validate, packages/sdk tests and the packages/web build all still pass.

[github-actions]

No actionable findings.