Qwen3.8-Omni-Flash - Tokenstead

Native omni model, API-only. Text, image, audio, and video in; text out on the standard Chat Completions / Responses API, with synthesized speech out on the realtime variant (WebSocket/WebRTC). Released 2026-09-14 by Alibaba’s Qwen team. Total parameters and architecture are undisclosed - the predecessor qwen3-omni-flash was a Thinker-Talker MoE, but Qwen has not published the 3.8 breakdown, so this card lists no parameter count.

• Context and I/O: 1M tokens in (991,808 usable without thinking), up to 131K out. Thinking is on by default with adjustable reasoning effort. Function calling, implicit caching, and Responses session caching are supported.
• Audio: ASR covers 113 languages and dialects (the same set as Qwen3.5-Omni); spatial (multichannel) audio input is supported. The realtime variant takes camera frames at 1 fps / 720p and handles about an hour of audio-visual input with speaker diarization.
• Benchmarks vs Gemini 3.8 Flash (Alibaba’s own numbers): wins the overall-audio and most audio-visual boards - WildClawBench-MM 71.0 vs 58.9, DailyOmni 85.1 vs 84.0, SpotSoundBench 67.2 vs 39.7, MMAU 81.8 vs 76.9, AliMeeting DER 3.4 vs 17.2 - and loses the agentic and video boards: OmniGAIA 74.0 vs 78.6, Video-MME-v2 65.0 vs 71.0, AgenticVBench 36.8 vs 45.0. Mixed, not a sweep - “beats Gemini on multimodal” is true on audio, not across the board.

No open weights. The model runs only via Alibaba Cloud Model Studio (Beijing and Singapore, plus Hong Kong, Tokyo, Frankfurt, and US-Virginia endpoints). Only the companion repos (Qwen-MM-Plugins, Qwen-Live-Harness) are open-sourced. The community reaction said it plainly: “No OSS :/”.

Cloud API: $0.15/1M input, $0.47/1M output, $0.016/1M cache-hit input on the Singapore endpoint ($0.113 / $0.382 / $0.014 on mainland/Global). Unlike its predecessor there is no per-modality split - audio and image tokens bill at the flat input rate, which is how Alibaba gets its “>98% cheaper per hour of audio than Qwen3.5-Omni-Plus” claim.