Qwen3.8-Omni-Flash

We track Qwen3.8-Omni-Flash, but the public leaderboard excludes this profile until enough non-generated benchmark coverage is available. Only published rows appear above.

Qwen3.8-Omni-Flash is a proprietary model with a 1M context window. It uses an explicit reasoning mode, which can improve complex problem solving while adding latency and token use.

Hosted only. Qwen serves qwen3.8-omni-flash through the Qianwen AI Platform and Alibaba Cloud Model Studio with OpenAI-compatible Chat Completions and Responses APIs, text, image, audio and video input, a 1,000,000-token context window, and reasoning_effort levels of low, medium, and xhigh (default). A separate streaming sibling, qwen3.8-omni-flash-realtime, is served over WebSocket and WebRTC for continuous audio-visual interaction; its throughput and latency figures are not attributed to this row. No weights are published.

Official exact-value snapshot from Qwen's September 18, 2026 Qwen3.8-Omni-Flash launch post. BenchLM maps only the text and vision table rows that match existing protocols. GPQA Diamond lands on both the weighted GPQA lane and the display GPQA-D lane, matching the existing provider-row convention; MathVision and CharXiv (RQ) keep their paired no-Code-Interpreter and Code-Interpreter values on separate keys; LiveCodeBench v6, DeepSWE 1.1, NL2Repo-Bench, and the in-house CoWorkBench stay display-only. HLE is stored on the weighted lane only, because Qwen labels the row as GPT-4o-judged without stating a no-tools setting. ClawEval-MM is reported as a pass@3 and average pair and stays out of schema, as on the Qwen3.8-Flash-Next row. The launch's 34-row omni table (WildClawBench-MM, UniClawBench, AgenticVBench, OmniGAIA, DailyOmni, WorldSense, AVUT, JoinAVBench, OmniVideoBench, Video-MME-v2, LVOmniBench, OmniCloze, OmniCap-IF, QIVD, StreamingBench, the four multi-speaker ASR sets, WenetSpeech, FLEURS ASR and S2TT, SpotSoundBench, MMAU, MMAR, MMSU, LongAudioSpan, the three music sets, Audio MultiChallenge, WildSpeech, and VoiceBench) has no equivalent in the weighted schema and is recorded as provider evidence on the voice-benchmark page instead of being collapsed into non-equivalent keys. The Agentic Omni Understanding table, which pairs static and Qwen Code agent runs on OmniVideoBench (63.4 to 67.8), Video-MME-v2 (65.0 to 71.3), and LVOmniBench (63.3 to 73.6), is a harness comparison and is recorded there too. The realtime throughput and latency table belongs to qwen3.8-omni-flash-realtime and is not stored on this row.

19 of 446 tracked benchmark slots currently have displayable evidence. Missing categories stay blank.

Its strongest eligible category is Instruction Following at #29, while its lowest eligible position is Knowledge at #54. a well-rounded choice across a range of tasks.