[ENHANCEMENT] Update Fireworks provider model list

[hannesrudolph]

Type

Enhancement

Problem / Value

The Fireworks provider’s built-in model list is out of date, which can cause users to pick models that no longer exist (or miss newer models that should be available).

Context

Affects anyone using the Fireworks provider in Roo Code. The provider’s model IDs and defaults are maintained in code and need a refresh to match Fireworks’ current catalog.

Constraints/Preferences

Keep backward compatibility where possible (don’t break existing configs; if a model is removed, ensure users can still see their currently selected model).

[roomote]

I'd like to help with updating the Fireworks provider model list. I've analyzed the current implementation in packages/types/src/providers/fireworks.ts and found the following models:

Current models in the codebase:

• Kimi K2 (kimi-k2-instruct, kimi-k2-instruct-0905, kimi-k2-thinking)
• MiniMax M2
• Qwen3 (qwen3-235b-a22b-instruct-2507, qwen3-coder-480b-a35b-instruct)
• DeepSeek (deepseek-r1-0528, deepseek-v3, deepseek-v3p1)
• GLM (glm-4p5, glm-4p5-air, glm-4p6)
• GPT-OSS (gpt-oss-20b, gpt-oss-120b)

To proceed with this update, I need some clarification:

1. Which models should be added? Could you provide a list of new Fireworks models that should be included?
2. Which models should be removed? Are any of the current models deprecated or no longer available on Fireworks?
3. Which models need updates? Are there any models with outdated pricing, context windows, or capabilities?
4. What should be the new default model? Currently it's kimi-k2-instruct-0905.

If you have access to Fireworks' current model catalog or can point me to documentation, that would help ensure the update is comprehensive and accurate.

Please mention @roomote in your response with these details, and I'll proceed with the implementation.

[lmtr0]

Fireworks has a openai compatible api, would you be able to add the list functionality like the openrouter / openai compatible pluggins work?

[mmealman]

I will confirm, that for GLM 4.7 at least, the OpenAI Compatible provider is working for me. Though a pain point for new users will be figuring out all the proper settings for each model. For example, attempting to use caching causes a failure.

Given how slow PR approvals have been for FireworksAI(and probably others), I wonder if it may make more sense to drop 80% of these extra providers and just focus on documentation on how to use OpenAI Compat with each provider and making sure OpenAI Compat works in general with the quirks on each provider.

[lmtr0]

It's probably misconfiguration on my part, but I can't get DeepSeek v3.2 to work on Roo Code with reasoning enabled