Problem#
Ten providers have activity: null on every row. These rows also have no modalities, no capabilities, and no rate card. #72 fixed the same gap for Groq and Mistral.
Measured on GET /v1/models?pricing=1 on 2026-10-03:
| Provider | Rows | activity set | Examples |
|---|
| dashscope | 172 | 0 | deepseek-v4-flash, kimi-k3 |
| novita | 121 | 0 | |
| perplexity | 51 | 0 | |
| jina | 30 | 0 | |
| fireworks | 29 | 0 | accounts/fireworks/models/deepseek-v4-pro |
| sambanova | 6 | 0 | DeepSeek-V3.2, gpt-oss-120b |
| hyperbolic | 5 | 0 | deepseek-ai/DeepSeek-R1 |
| cerebras | 2 | 0 | gpt-oss-120b, qwen-3.8-27b |
| deepseek | 2 | 0 | deepseek-flash, deepseek-v4-pro |
| moonshot | 2 | 0 | kimi-k2.6, kimi-k2.7-code |
Together is a related case. Its 187 chat rows have activity: "chat", but modalities and capabilities are null on all of them (example: agentica-org/DeepCoder-14B-Preview).
Impact#
?activity=chat returns nothing for these providers. TanStack AI's @tanstack/ai-models generator cannot tell a chat model from an embedding model, cannot read input modalities, and gets no price. It cannot move these providers from models.dev to modelschemas.
Ask#
- Fill
activity and schemaEndpointId for the ten providers, the same way as #72.
- Fill
modalities and capabilities for these providers and for the Together chat rows, from a native source only.
- Add a pricing source for DeepSeek, Cerebras, Moonshot, and Fireworks first. They are the ones TanStack AI reads.
Acceptance#
GET /v1/models?provider=deepseek&activity=chat returns the chat models. The same is true for cerebras, moonshot, and fireworks.
GET /v1/status shows a non-zero counts.priced for deepseek, cerebras, moonshot, and fireworks.
- Together chat rows have
modalities set.