[jimmyplup]

Adds Synap by Linkrra (OpenAI-compatible) as a provider, with mappings for the three models we serve only on GPUs we operate. Opened at the maintainer's invitation (email thread with @umid, 20 Oct).

Public sources

• API base URL: https://pool.linkrra.com/v1 (bearer auth, OpenAI-compatible: streaming, tools, response_format)
• Prices, served context limits and supported parameters: https://pool.linkrra.com/v1/models (public, no login) — also https://synap.linkrra.com/models
• Live serving nodes per model: https://pool.linkrra.com/v1/nodes (public)
• Docs: https://synap.linkrra.com/docs
• Sign-up (free tier, no card): https://synap.linkrra.com/account
• Plans and rate limits: https://synap.linkrra.com/pricing
• Privacy / data handling: https://synap.linkrra.com/privacy

Mappings (values read from the live feed, 2026-10-21)

model_id	API model id	in / out per 1M (USD)	served context	capabilities in the feed
gpt-oss-20b (existing)	gpt-oss-20b	0.017 / 0.085	32,768	tools, tool_choice none/auto/required, structured outputs (json_object, json_schema)
gemma-4-26b (existing)	gemma-4-26b-a4b-it	0.04 / 0.21	131,072	tools, tool_choice none/auto/required, structured outputs (json_object, json_schema); text input only on our deployment
qwen3-coder-30b-a3b (new)	qwen3-coder-30b	0.0675 / 0.27	32,768	tools, tool_choice none/auto/required, structured outputs (json_object, json_schema)

• New canonical model qwen3-coder-30b-a3b: the existing qwen3-coder entry does not name a variant (its Together mapping is Qwen/Qwen3-Coder), so I did not reuse it. Values from the model card https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct (30.5B total / 3.3B active, 262,144 native, 1M with YaRN, Apache-2.0); context_window: 1000000 following the convention in your review of Add Maison Token provider and Qwen3.8-27B mapping #7, with our served limit as context_window_override.
• Weights served: gpt-oss-20b in its native MXFP4; Qwen3-Coder-30B-A3B and Gemma 4 26B A4B as AWQ 4-bit quantizations. Flagging this in case you want quantized deployments described differently.
• gemma-4-26b: reused because it is the only Gemma 4 26B entry and our weights are google/gemma-4-26B-A4B-it. Note the canonical description calls it a dense model with all parameters active; Google's card describes 26B A4B as the mixture-of-experts variant. Happy to fix that in this PR or a separate one if you prefer.
• Rate limits are per plan (20 to 1,000 requests/min, https://synap.linkrra.com/pricing), with no published TPM, so both are null.
• Region: global is the API's reach. The datacenters field in the feed currently shows the serving nodes in South Korea (KR); they are rented GPUs and the location can change.

Routing and data handling — why only these three

• These three models have no third-party route. No other provider serves them: they are keyed by our own model ids, which no upstream recognises. With no node of ours online the request returns an error, not an answer from elsewhere. One caveat for completeness: in that state our gateway currently still forwards the request to OpenRouter, under the same zdr: true / data_collection: deny flags, and OpenRouter rejects the unknown id. We are changing that so it fails inside our gateway.
• Not included on purpose: qwen/qwen3-30b-a3b and qwen/qwen3-vl-30b-a3b-instruct are also served on our nodes, but when none is online they fall back to a third-party provider through OpenRouter (zero-data-retention endpoints only: zdr: true, data_collection: deny, set per request in our code), billed at a higher brokered price. One price field can't describe that, so they are left out.
• Brokered models in our feed (pricing.basis: brokered, e.g. DeepSeek, GLM, Kimi) and our router/team aliases are not in this PR.
• Why qwen3-coder-30b showed is_ready: false / basis: brokered earlier: the feed quotes the self-hosted price only while one of our nodes for that model is online. While it was offline the feed fell back to the catalogue's default price basis. That label is about the price, not the route; the model has no upstream route.

Validation

.venv/bin/python validate_registry.py
✅ Validation Passed
Validated 569 files successfully.

No API keys in this PR.