Wait Which Model?
← Back to directory
Alibaba (Qwen)

Qwen3.8-Omni-Flash

Released Sep 18, 2026 · knowledge cutoff unpublished

Status
Unknown
Location
China
Modality
Multimodal
Context window
1M
Max output
131K
Speed
Price ($/MTok in / out)
$0.11 / $0.38
Cost per task
Open weights
No
Benchmarks2 of 10 reported

GPQA Diamond

91%

SWE-bench Pro

63.3%

MMLU-Proretired

SWE-bench Verified

Terminal-Bench 2.1

AIME (latest)retired

Humanity's Last Exam

LMArena Elo

GDPval-AA v2

ARC-AGI-2

Strengths & weaknesses

Strengths

  • Reasons over audio and video with an agentic, coarse-to-fine 'selective seeking' pass — decides what to watch or listen to first and spends tokens only on the segments that matter, rather than transcoding a whole stream
  • Alibaba's own eval of real meeting audio shows diarization errors falling from 88% to 3% and transcription errors from near 90% to 17% versus the prior Omni line
  • Sharply cheaper on audio/video workloads than its predecessor — Alibaba reports a 98% drop in per-hour audio cost and 93% on combined audio-video

Weaknesses

  • Text-only output — no generated speech; Alibaba's own Model Studio docs point developers back to Qwen3.5-Omni when a spoken response is needed
  • Selective seeking trims tokens at the cost of more tool round-trips, decode time and storage reads, so latency doesn't fall as cleanly as the token count does
  • Hosted-only despite sitting on the open-weight Qwen3.8-Flash-Next architecture — Alibaba has published no downloadable weights for the Omni-Flash checkpoint itself

Alibaba's first omni-modal model built around agentic tool use (text, image, audio and video in; text out), launched 2026-09-18 on Qwen Chat, QwenCloud and Model Studio's OpenAI-compatible endpoint. GPQA Diamond 91.0 and SWE-bench Pro 63.3 are Alibaba's own reported figures via launch coverage (official qwen.ai/alibabacloud.com pages were unreachable this session); MMLU-Pro, Terminal-Bench, AIME, HLE, LMArena Elo, GDPval-AA and ARC-AGI-2 were not found reported for this specific checkpoint and are logged to stats-gaps.md. Pricing (~$0.11/$0.38 per MTok) is converted from Alibaba Cloud Model Studio's official CNY rate (0.8/2.7 per million tokens) and matches one reseller's quote, but a second reseller lists $0.15/$0.47 — flagging the conflict rather than picking one; per-hour audio/video surcharge pricing is not broken out here. Context window is Alibaba's marketed 1M figure; QwenCloud's own listing shows a 991K/131K input/output split, used for maxOutput. Coverage repeatedly calls Qwen3.5-Omni-Plus this model's 'predecessor' by name, but Qwen3.5-Omni-Plus is not itself a tracked model in this dataset, so predecessorId stays null (logged to spec-gaps.md). No Artificial Analysis coverage found, hence costPerTask and speed are null; no official retirement page exists yet.

For developers

API model strings

  • Alibaba Cloud Model Studioqwen3.8-omni-flash

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News