Qwen3.5 122B A10B
Intelligence
16.19
#250 of 520 models
Output speed
127.9 tok/s
Measured generation rate
Cost / index task
$0.325
Measured evaluation workload
End-to-end latency
22ms
First answer 18ms
01 / Position
A model in the landscape
Compare this model with peers across intelligence, speed, cost, and individual evaluations.
02 / Evidence
Benchmark results
24 distinct evaluations grouped by topic. Explore cross-model charts or switch to the full results list.
Domain capability
03 / Operation
Speed, tokens & price
Prices are per million tokens; latency and task costs describe the captured evaluation rather than a universal provider guarantee.
Token pricing
- Input / 1M
- $0.400
- Output / 1M
- $3.20
- Cache read / 1M
- —
- Cache write / 1M
- —
- Cache-hit discount
- —
Response & usage
- Output speed
- 127.9 tok/s
- First chunk
- 2ms
- First answer
- 18ms
- Output / task
- 17.1k
- Reasoning / task
- 7.3k
Evaluation workload
- Canonical input
- 847.29M
- Canonical output
- 100.74M
- Answer / task
- 9.8k
- Index time / task
- 2.1m
- Run cost
- $661.288
By prompt type
| Prompt | Output speed | First chunk | First answer | End to end |
|---|---|---|---|---|
| Medium | 134.1 tok/s | 1ms | 16ms | 20ms |
| Long | 127.9 tok/s | 2ms | 18ms | 22ms |
| Hundred K | 106.7 tok/s | 20ms | 38ms | 43ms |
| Medium Parallel | 121.8 tok/s | 1ms | 17ms | 22ms |
04 / Access
License & availability
Recorded license
Apache 2.0
The publisher-linked license text defines permitted use and redistribution. Review it before using the model.
- Weight access
- Open weights
- Classification
- Permissive
- Reasoning
- Yes
Routing and providers
- OpenRouter ID
- qwen/qwen3.5-122b-a10b
- Tokenizer
- Qwen3
- Instruct type
- —
- Captured endpoints
- 5
| Provider | Quantization | Context | tok/s |
|---|---|---|---|
| SiliconFlow | fp8 | 262.1k | — |
| Alibaba | unknown | 262.1k | — |
| DeepInfra | fp4 | 262.1k | — |
| AtlasCloud | fp8 | 262.1k | — |
| Novita | bf16 | 262.1k | — |
05 / Audit
Underlying fields
Inspect the captured model and routing records. Artificial Analysis destination links are omitted.