Qwen3.8-Omni-Flash for audio agents: tool access

The Qwen release feed for September 17, 2026 lists Qwen3.8-Omni-Flash, described as native omnimodal across audio, vision and text for agentic productivity and creative work. The feed says nothing about Sume. Any agent on any model can use Sume's hosted MCP at https://mcp.sume.com/mcp if its client speaks remote HTTP MCP.

What the release says

Per the Qwen release updates page, read on 2026-10-03, Qwen3.8-Omni-Flash is native omnimodal (audio, vision, text). That is the full extent of what this post asserts about the model. Check Qwen's own documentation for limits, pricing and tool-calling support.

What Sume exposes to an agent

Hosted MCP wraps selected public API capabilities as tools. Paid generation tools include generate_image, generate_video, music_create, tts_create and stt_create. Job tools include jobs_status, jobs_wait and jobs_result. It is not full parity with the HTTP API, so discover the live contract with tools_list and tools_schema.

Gates that apply to any model

Paid and write tools require an idempotency_key. dry_run=true previews cost without submitting, and max_spend_usd is enforced only when provided. Under OAuth the default scope is read-only; write and paid tools stay hidden until mcp:write is granted or an API key is used.

{
  "idempotency_key": "voiceover-take-001",
  "dry_run": true,
  "max_spend_usd": 1
}

A sensible first run

Connect the client, call mcp_health, then tools_list. Ask the agent to restate which tools it can see before it plans. If the model cannot call tools reliably, keep it on read-only discovery and put a human between it and any paid call.