Parameters: ollama-cloud/beam #424

Model: ollama-cloud/beam
Auth type: Subscription

Request: add configurable parameters for Beam (Reflection 501B-A23B MoE, served via Ollama Cloud)

Beam is currently served with no configurable model params on the routing
params API (mnfst routing params get returns an empty params list for the
tier/model), which makes it unusable as a gate seat for non-trivial inputs.

The problem we hit in production (3 real dispatches, request ledger)

Beam is a thinking model with a very verbose reasoning style that shares
the completion-token budget between reasoning and visible content. With real
review-gate payloads (~68–78K input tokens):

Ledger shows status: success, finish_reason: stop — the failure is
invisible to the gateway because the model did finish; it just spent the
whole output budget on thinking and produced minimal visible content.

A controlled streaming probe with the same payload shape shows the split:

• completion_tokens: 2210 total
• reasoning (delta.reasoning): ~8,900 chars (~90% of the budget)
• visible content: ~640 chars (~150 tokens)

The reasoning style is verbose planner text ("We need answer code review
findings. Need output findings first…"). With ~52K-token prompts, the
non-streaming path can also exceed the gateway/provider patience (observed a
196s hop before the fallback served the request).

What we need

Any of these would make the model viable as a routed seat:

1. reasoning_effort / thinking control — equivalent to the think /
   reasoning toggle other Ollama Cloud models expose, so reasoning can be
   reduced or disabled per request, and set per tier/model via
   mnfst routing params set (v6.26.0 params feature).
2. max_tokens (or a separate reasoning-token budget) — registered as a
   configurable param so the tier can raise the completion ceiling enough for
   reasoning + content to coexist (we'd set ≥16384).
3. Ideally both, so effort can be capped and the budget sized.

Context

• Served id on our gateway: ollama-cloud/beam (Ollama Cloud subscription;
  upstream free-promotion endpoint, ZDR).
• API mode: chat_completions, OpenAI-compatible translation path.
• Verified live: small payloads work fine (Say OK. → 49 in / 43 out,
  finish stop); the failure mode only appears with large inputs where thinking
  dominates the budget.
• Without params control, we had to demote the model from a primary seat to
  fallback-only, despite a benchmark profile (reasoning/tool-use) that would
  otherwise fit gate work.