feat(tium): add Tium provider with the GLM, DeepSeek and Kimi families

[curiousbox]

Adds Tium, an OpenAI-compatible gateway operated in Germany that resells open-weight models from Z.ai, DeepSeek and Moonshot. Five models, all probed live before authoring. Entries will be kept current.

Provider

• providers/tium/provider.toml: npm = "@ai-sdk/openai-compatible", api = "https://api.tium.ai/v1", env = ["TIUM_API_KEY"], doc = "https://tium.ai/models".
• providers/tium/logo.svg: single path, fill="currentColor", square viewBox="0 0 32 32", no width or height attributes.
• Every hand-authored file carries a leading comment naming its sources and the date the values were verified.

Models

All five use base_model against lab entries that already exist in the catalog, and author only the provider-side fields. No models/ files are added or modified.

provider file	base_model	input	output	reasoning	cache_read	reasoning_options
glm-5.3-flash	zhipuai/glm-5.3-flash	0.152	0.505	0.505	0.03	effort: low, high, max
deepseek-v4-flash	deepseek/deepseek-v4-flash	0.889	2.668	2.668	0.028	toggle + effort: low, high, max
deepseek-v4-pro	deepseek/deepseek-v4-pro	2.668	8.005	8.005	0.089	toggle + effort: high, max
glm-5.3	zhipuai/glm-5.3	2.83	8.895	8.895	0.526	effort: low, high, max
kimi-k3	moonshotai/kimi-k3	6.065	30.323	30.323	0.606	toggle + effort: low, high, max

Each also sets limit.context = 128000, limit.output = 32768 and [interleaved] field = "reasoning_content".

Pricing basis

Costs are USD per million tokens, as required. Tium does not bill USD per token: it sells a monthly allotment of weighted tokens (WT). The figures above are the effective USD rate at the subscription price of $2.8302/MWT, the highest of the three published tiers, multiplied by each model's published multiplier and its output and cache weights. Prepaid credit packs bill higher, $3.56 to $4.50/MWT, so the published rate is the conservative end. Reasoning tokens are counted and billed as output tokens, hence cost.reasoning matching cost.output. Source: https://tium.ai/pricing. There is no context-based tiering on this host, so no [[cost.tiers]].

limit.context and limit.output are this host's policy ceilings, not the base models' maximums, which are higher. They are authored on the provider files for that reason.

Verification

Probed live against api.tium.ai on 2026-09-09. Per-file comments carry the measurements.

• Reasoning controls were tested, not assumed. glm-5.3 and glm-5.3-flash accept reasoning_effort of low, high and max; none, minimal, medium and xhigh return 400, so the negative control fired. thinking.type = disabled is also refused, making both models thinking-only on this host, which is why neither publishes a toggle.
• Effort levels are published only where they proved distinguishable. For glm-5.3 and glm-5.3-flash, reasoning_content medians over 3 runs on a fixed prompt were 94/266/753 and 121/252/1241 characters, with non-overlapping ranges. For the DeepSeek and Kimi models the wire accepts minimal, medium and xhigh, but medians across those levels overlapped with no ordering, and they are absent from the lab baselines, so they are omitted rather than published as real controls. The per-file comments record this and note that a 600-token output ceiling on the measurement compresses differences at the top end.
• Toggles were confirmed to be real. On the three models that publish { type = "toggle" }, thinking.type = disabled genuinely removes reasoning_content from the response rather than merely being accepted.
• Modalities were narrowed to what this host accepts, rather than inherited optimistically. kimi-k3 rejects video with "unsupported video url" despite the lab entry listing it, and refuses PDF, hence the override to text and image. glm-5.3-flash likewise overrides to text and image. glm-5.3 rejects image, PDF and video alike, which matches its lab entry, so it carries no override. The DeepSeek models accept image and PDF parts, but the lab entries are text-only and acceptance alone does not demonstrate comprehension, so no extra modality is claimed.

Notes

• Three routes are served: GET /v1/models, POST /v1/chat/completions and POST /v1/completions. POST /v1/responses returns 404, so clients that default to the Responses API need the chat-completions wire format. This is recorded in the provider comment.
• No sync module is proposed. GET /v1/models requires an API key, so it cannot be polled unauthenticated the way sync.md modules expect, and a five-model catalog is below the threshold where one earns its keep. Happy to add one if you would prefer it.

[github-actions]

No actionable findings.