[jaberjaber23]

Summary

Adds RunInfra (https://runinfra.ai) as an OpenAI-compatible inference provider so Crush users can select it directly.

RunInfra serves open-weight models behind an OpenAI-compatible endpoint at https://api.runinfra.ai/v1, authenticated with RUNINFRA_API_KEY. This adds the provider config (internal/providers/configs/runinfra.json) plus the matching pkg/catwalk constant and internal/providers registry entry, so go test ./... passes.

Models

Model	id	Context	Image input	Input $/1M	Cached $/1M	Output $/1M
DeepSeek V4.1 Flash	deepseek-v4-1-flash	1,048,576	No	0.14	0.03	0.58
GLM 5.3 Flash	glm-5-3-flash	1,048,576	Yes	0.11	0.03	0.45
Qwen3.8 27B	qwen3-8-27b	262,144	Yes	0.10	0.01	0.40
Ornith 1.5 35B	ornith-1-5-35b	262,144	Yes	0.10	0.01	0.40
Nemotron 3.5 Lightning 30B	nemotron-3-5-lightning-30b	262,144	No	0.05	0.01	0.15

Defaults: large deepseek-v4-1-flash, small nemotron-3-5-lightning-30b.

Notes

• Prices, context windows, and image-input support were taken from the public per-model pages (https://runinfra.ai/inference-api/<id>) and the pricing page, not from memory. Only deepseek-v4-1-flash and nemotron-3-5-lightning-30b are text-only on this serve; the other three accept image input.
• The single advertised cached-input rate is placed in cost_per_1m_out_cached (cache read) with cost_per_1m_in_cached left at 0, following the convention documented in CRUSH.md.
• The model id is the slug RunInfra's own code examples pass as the model parameter (e.g. deepseek-v4-1-flash).
• Verified locally with go build ./... and go test ./...; the internal/providers tests confirm the provider is registered in KnownProviders() and that the default model ids exist.

Source: https://runinfra.ai/pricing

🤖 Generated with Claude Code