Qwen3.8-27B
Released Aug 14, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- China
- Modality
- Multimodal
- Context window
- 262K
- Max output
- 131K
- Speed
- —
- Price ($/MTok in / out)
- $0.45 / $3.20
- Cost per task
- —
- Open weights
- Yes
Benchmarks2 of 8 reported
GPQA Diamond
Terminal-Bench 2.1
MMLU-Pro
SWE-bench Verified
AIME (latest)
Humanity's Last Exam
LMArena Elo
ARC-AGI-2
Strengths & weaknesses
Strengths
- Runs comfortably on a single high-end consumer GPU while staying competitive with Alibaba's own cloud-only siblings on coding and office tasks
- Vision and video understanding are built into the same dense weights from the start, not bolted on as a separate adapter
- Reasoning effort is adjustable at inference time (off through xhigh), so a quick lookup and a long agent run can share one download
Weaknesses
- No competition-math (AIME) figure has been published, so its strength relative to peers on hard math is undocumented
- Trails the closed frontier flagships on the hardest science and knowledge questions — a 27.8B dense model has a real ceiling
- No technical report or knowledge cutoff disclosed at launch, so training-data freshness and provenance are unclear
27.8B dense (not MoE) multimodal model — distinct from the 2.4T-parameter MoE flagship qwen3-8-2-4t-a95b (2026-08-13) and the closed qwen3-8-max, and is the smaller open checkpoint Alibaba promised alongside Qwen3.8-Max's 2026-08-03 GA announcement. GPQA Diamond 89.2% and Terminal-Bench 2.1 73.0% (up from 63.4% for Qwen3.6-27B) are from Alibaba's own release material; SWE-bench Pro 61.7% and OSWorld-Verified 84.3% were also reported but are not the SWE-bench Verified variant this site tracks. Qwen's own announcement (X/@Alibaba_Qwen) says it 'outperforms Qwen3.7-Plus overall' — comparison language, not a replacement statement, so predecessorId is left null. Pricing is OpenRouter's hosted rate ($0.45/$3.20 per 1M in/out); Alibaba has not published a first-party DashScope price for this open-weight checkpoint. Context is 262,144 native, extendable to roughly 1M via YaRN; max output is the 131,072 cap.
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor