Fireworks AI reference for Nexus BYOC

Nexus BYOC is available only on Enterprise plans.

This page lists the Fireworks models that work with Pinecone Nexus Bring Your Own Cloud (BYOC) and the inference-proxy-models.toml fields a Fireworks catalog entry uses. For setup steps, see Use Fireworks AI models in Nexus BYOC.

Supported models

The following serverless models work with Nexus BYOC. For the full list, including models you can run on dedicated deployments, see the Fireworks model library.

Model type	Catalog ID	Fireworks model path	Suggested use
Chat	`glm-5p3-flash`	`accounts/fireworks/models/glm-5p3-flash`	Curation and System defaults.
Chat	`deepseek-v4p1-flash`	`accounts/fireworks/models/deepseek-v4p1-flash`	Curation default.
Chat	`glm-5p3`	`accounts/fireworks/models/glm-5p3`	Selected per query.
Chat	`kimi-k3`	`accounts/fireworks/models/kimi-k3`	Query default, for complex, multi-step questions.
Embedding	`qwen3-embedding-8b`	`accounts/fireworks/models/qwen3-embedding-8b`	Optional default embedding model. Returns 4,096-dimensional vectors.

You can also use accounts/fireworks/models/gpt-oss-120b and accounts/fireworks/models/minimax-m3p1-flash the same way. Set each entry’s context_window to the model’s own window, such as 131_072 for gpt-oss-120b. gpt-oss-120b is text-only, so also set no_vision = true. The catalog ID is the name you choose for an entry, and the examples here match it to the model name.

Endpoint and authentication

Endpoint: Set base_url to https://api.fireworks.ai/inference/v1, the Fireworks OpenAI-compatible endpoint.
API key: Store your Fireworks API key as the stack secret nexus-provider-keys.<ref>, and set the entry’s api_key_ref to <ref>. Every Fireworks entry can share one reference. Pulumi stores the secret as a Kubernetes secret, and the inference proxy reads it at startup. If the key doesn’t resolve, the proxy lists the model as unavailable.

Chat model fields

A Fireworks chat model is an [llm_models.<id>] entry with api_style = "openai".

Field	Required	Description
`api_style`	Yes	Set to `openai`.
`model`	Yes	The Fireworks model path, such as `accounts/fireworks/models/glm-5p3`.
`base_url`	Yes	The Fireworks endpoint.
`api_key_ref`	Yes	The name of the secret that holds your Fireworks API key.
`label`	Yes	The model name shown in the console and in `GET /models`.
`provider`	Yes	The provider tag shown in the console and in `GET /models`, such as `fireworks`.
`context_window`	Yes	The input token budget.
`max_output_tokens`	Yes	The output token cap. The proxy enforces it on every request.
`input_price_per_mtok`	No	The price per million input tokens. Set the price fields to have usage and the query trace report cost for the model.
`output_price_per_mtok`	No	The price per million output tokens.
`cache_read_price_per_mtok`	No	The price per million cached input tokens.
`cache_write_price_per_mtok`	No	The price per million tokens written to the cache.
`no_vision`	No	Set to `true` for a model that doesn’t accept images. Defaults to `false`, which advertises the model as vision-capable.

For current prices, see Fireworks pricing.

Embedding model fields

A Fireworks embedding model is an [embedding_models.<id>] entry with api_style = "openai".

Field	Required	Description
`api_style`	Yes	Set to `openai`.
`model`	Yes	The Fireworks model path, such as `accounts/fireworks/models/qwen3-embedding-8b`.
`base_url`	Yes	The Fireworks endpoint.
`api_key_ref`	Yes	The name of the secret that holds your Fireworks API key.
`dimension`	Yes	The width of the vectors the model returns. Set it to the model’s actual output width.
`max_input_chars`	Yes	The length each input is truncated to before the call.
`max_batch_size`	Yes	The most inputs per upstream call. The proxy splits larger requests into batches of this size.
`provider`	No	The provider tag used to split spend by vendor, such as `fireworks`.

Default profile fields

The [default] profile controls which catalog models your deployment serves.

Field	Description
`supported_llm_models`	The chat models the deployment serves. Every model a default or a query names must be in this list.
`supported_embedding_models`	The embedding models the deployment serves.
`supported_rerank_models`	The rerank models the deployment serves.
`embedding.tiers.default`	The default embedding model, set with `model_ref`. Required.
`rerank.tiers.default`	The default rerank model, set with `model_ref`. Required.
`llm.phase_defaults`	The default model for each work area: `search` for Query, `curate` for Curation, and `agent` for System.
`llm.tiers`	The `lite`, `standard`, and `pro` tiers, each set with `model_ref`. A work area that `phase_defaults` doesn’t set falls back to a tier.

Constraints

Chat model parameters: Every chat model must support the OpenAI tools and response_format parameters, because Nexus relies on both.
Default models: Each work area’s default comes from llm.phase_defaults, falling back to the tiers: Query from search or the standard tier, Curation from curate or the lite tier, and System from agent or the standard tier. The catalog reapplies these defaults on every settings pull, so the console’s Default models card is read-only. To change a default, edit the catalog and run pulumi up.
Tier names in queries: A query treats a tier name, such as pro, as the Query default. Name models by catalog ID instead.
Embedding changes: A Pinecone index’s dimension is fixed when the index is created. If you change the default embedding model, keep every embedding model a context has used in the catalog, because existing contexts look up their embedding model by ID at query time.

Get started

The knowledge engine

Tools

Bring Your Own Cloud

Fireworks AI reference for Nexus BYOC

Supported models

Endpoint and authentication

Chat model fields

Embedding model fields

Default profile fields

Constraints

​Supported models

​Endpoint and authentication

​Chat model fields

​Embedding model fields

​Default profile fields

​Constraints

Supported models

Endpoint and authentication

Chat model fields

Embedding model fields

Default profile fields

Constraints