brahma-30b-reasoning is your organization’s self-hosted reasoning model. Before it answers, it works through the problem internally — planning, checking, and revising — then returns a clean final answer. Use it for multi-step math, code, analysis, and any task where a balanced chat reply falls short. It speaks the same Chat Completions API as brahma-30b. To use it, send "model": "brahma-30b-reasoning" on the same endpoint — no new base URL, no new client.

​
The two tiers

Both run on your own GPUs and share the same identity, languages, tools, and structured-output support. Start with brahma-30b; reach for brahma-30b-reasoning when a task needs deliberate, multi-step thinking.

​
How the reasoning is handled

The model produces a private reasoning trace and then the final answer. Like other frontier reasoning APIs, the raw chain-of-thought is not returned — message.content contains only the final answer, so you can render it directly. The thinking is still metered: usage.completion_tokens_details.reasoning_tokens reports how many tokens were spent thinking, and those tokens are already included in completion_tokens (they bill as output). This is the only shape difference from a brahma-30b response.

​
Controlling how hard it thinks

Pass reasoning_effort to trade latency and cost against depth: Higher effort means more reasoning tokens (and a longer response), so high costs more per call. The field is optional; omit it for the model default. It has no effect on brahma-30b (accepted and ignored), so shared client code stays simple.

​
Example

​
Prompting tips

  • Don’t ask it to “think step by step.” It already does — extra instructions to reason aloud only fights the built-in trace. State the task and the answer format you want.
  • Give it the goal, not the procedure. Reasoning models do best when you describe what a good answer looks like and let the model find the path.
  • Use high sparingly. Most tasks are well served by the default; reserve high for problems that genuinely need it, since it spends the most reasoning tokens.
  • Streaming returns only the final answer’s tokens; the thinking is not streamed. The final usage frame carries reasoning_tokens.
The reasoning plane runs on its own scale-to-zero GPU: the first request after an idle period takes a few extra seconds while a worker warms. Because the model thinks before answering, expect a longer time-to-first-token than brahma-30b — this is the model reasoning, not a fault.