Open-weight experimental preview of the Qwen4 architecture: hybrid-attention MoE (125B total, 6B active) with vision encoder for coding, agent tasks, and image and video understanding
Specification
most-agreed values
Context
262K
Max output
131K
Released
2026-08-27
Knowledge cutoff
—
Retires
—
Open weights
yes
Input
text, image, video
Output
text
Price
US dollars per million tokens · most-agreed
Input
$0.10
Output
$0.20
Cache read
$0.05
Cache write
$0.20
Quality
1 benchmark · 1 source
Composite
74th
percentile of 344 scored models
Rank
89 / 344
±0.56 sd
Evidence
1 × 1
too little evidence to place with confidence
Effort range
—
one configuration only
general59th1/1 bench
Every score
one row per source and configuration — nothing averaged away
| Benchmark | Score | Configuration | Source | Run |
|---|---|---|---|---|
| LiveBench | 77.3 | default · 23 runs | LiveBench | 2026-06-25 |
The leading figure for a benchmark is the median across its configurations, so one heroic high-effort run cannot set the number.
Available from 5 hosts
Price history
append-only observations · a listing writes a row only when its price moves
| Date | Host | Field | $/M |
|---|---|---|---|
| 2026-09-02 | amd | Cache read | $0.016 |
| 2026-10-22 | amd | Cache read | free |
| 2026-09-02 | amd | Input | $0.15 |
| 2026-10-22 | amd | Input | free |
| 2026-09-02 | amd | Output | $0.47 |
| 2026-10-22 | amd | Output | free |
More from Alibaba
most-hosted first