[BUG] Thinking mode for DeepSeek V3.1 on Fireworks AI provider is not working.

[kavehsfv]

Problem (one or two sentences)

The “Thinking Mode” (reasoning mode) for DeepSeek-V3.1 is not functioning on the Fireworks AI provider. It means when I run prompt, thinking section is not appear. If I select Minimax m2 model, the thinking section is appear and works properly.

Context (who is affected and when)

everyone uses DeepSeek-V3.1 on fireworks ai

Reproduction steps

select DeepSeek-V3.1 on Fireworks ai provider

Expected result

adding thinking feature for this model

Actual result

.

Variations tried (optional)

No response

App Version

4.137.0

API Provider (optional)

Fireworks AI

Model Used (optional)

DeepSeek-V3.1

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I'd like to help with this issue, but I need some clarification to ensure I implement the right solution. I've analyzed the codebase and found the following:

Technical Background:

• DeepSeek V3.1 (deepseek-v3p1) on Fireworks is currently configured as a non-reasoning model
• DeepSeek R1 (deepseek-r1-0528) is the reasoning variant that natively outputs thinking traces
• The codebase handles reasoning in two ways:
  1. Via <think> tags in the response (parsed by XmlMatcher)
  2. Via dedicated reasoning_content or reasoning fields in the API response

Clarifying Questions:

1. Model expectation: Are you expecting DeepSeek V3.1 to output thinking/reasoning traces similar to DeepSeek R1? Note that V3.1 is described as having 'better reasoning capabilities' but is not designed to output visible thinking traces like R1 models.

2. MiniMax M2 comparison: When you mention MiniMax M2 shows thinking sections correctly, are you using it through:

   • The dedicated MiniMax provider (which has native Anthropic-style thinking support), or
   • The Fireworks provider with the accounts/fireworks/models/minimax-m2 model?

3. Fireworks API: Do you have confirmation from Fireworks documentation that DeepSeek V3.1 supports a thinking/reasoning mode via API parameter? If so, we would need to enable the supportsReasoningBinary flag for this model.

4. Alternative: Would using deepseek-r1-0528 on Fireworks meet your needs instead? This is the reasoning model variant that outputs thinking traces.

Please mention @roomote in your response with answers, and I'll be happy to implement the appropriate fix once I have this information.

[kavehsfv]

@roomote

1. Yes I expect DeepSeek V3.1 has output thinking/reasoning traces similar to DeepSeek R1, as it shows reasoning on their fireworks website when I use it.
2. Fireworks provider with the accounts/fireworks/models/minimax-m2 model
3. Yes, Fireworks documentation confirms that DeepSeek V3.1 supports a thinking/reasoning mode via API parameter
4. Sure

[roomote]

Thanks for confirming! I'll work on enabling thinking mode support for DeepSeek V3.1 on the Fireworks provider. Since you confirmed that Fireworks documentation supports the thinking/reasoning mode via API parameter, I'll add the supportsReasoningBinary flag for the deepseek-v3p1 model in the Fireworks configuration.

I'll open a PR shortly with this fix.

[roomote]

Opened a PR with my attempt at addressing this: #10072

The fix adds supportsReasoningBinary: true to the DeepSeek V3.1 model configuration on Fireworks. When you enable reasoning in the extension settings, this will pass the thinking: { type: "enabled" } parameter to the Fireworks API, which should trigger the reasoning/thinking mode.

Feedback welcome!

[kavehsfv]

@roomote
Could you please add DeepSeek V3.2 to the Fireworks AI provider as well?

[roomote]

I've seen your request to add DeepSeek V3.2 to the Fireworks AI provider. I'll work on adding this model to the Fireworks configuration. I'll open a PR shortly with this addition.

[roomote]

Opened a PR with my attempt at adding DeepSeek V3.2 to the Fireworks AI provider: #10076

The PR adds the accounts/fireworks/models/deepseek-v3p2 model with configuration similar to V3.1. Feedback welcome!