Glowing Spoon

Industry / Models / Qwen / Alibaba

Qwen

A family spanning hosted reasoning models and open-weight coding checkpoints for configurable deployment.

Model snapshot
Qwen3.8-Max benchmarks; Qwen3-Coder open-weight option
Deployment
Hosted API; selected checkpoints can be served with compatible inference stacks
Context & outputs
Qwen3-Coder-480B-A35B: 262,144 tokens natively
Access & license
Qwen3-Coder checkpoint: Apache 2.0; hosted model terms are separate

Hosted and open-weight options

Qwen3.8-Max provides a hosted benchmark reference across coding, scientific reasoning, and instruction following. Those results describe that model only. Qwen3.8-Max release.

For deployment control, Qwen3-Coder-480B-A35B-Instruct is an Apache-2.0 checkpoint with 480B total parameters and 35B activated per token. Its native context is 262,144 tokens; the published extension to 1M uses YaRN and should not be treated as the native configuration. Official checkpoint.

Coding and tool use

The coding family was developed for repository-scale and tool-driven tasks, including a dedicated function-call format and Qwen Code integration. Qwen3-Coder technical overview.

Our engineering approach

We choose the checkpoint and serving configuration together. For self-hosted work, we measure GPU memory, quantization effects, concurrent throughput, and tool-call correctness. We keep hosted Max scores separate from open-weight Coder results, then run the same acceptance suite on the actual deployment candidate.

Published evidence

Benchmark snapshot.

Scores below are for the hosted Qwen3.8-Max release, not Qwen3-Coder. The open-weight checkpoint has a different architecture, release date, and evaluation profile. Values are Qwen-reported.

Published results for Qwen3.8-Max benchmarks; Qwen3-Coder open-weight option; higher is better within each metric. Results have different scopes and are not a universal ranking.
Parameter / benchmarkResultEvaluation settingSource
Software engineeringSWE-bench Pro67.7%Qwen3.8-Max; provider-reportedPublished result
Terminal agentsTerminal Bench 2.186.6%Qwen3.8-Max; provider-reportedPublished result
Scientific reasoningGPQA Diamond92.6%Qwen3.8-Max; provider-reportedPublished result
Instruction followingIFBench82.8%Qwen3.8-Max; provider-reportedPublished result

These are published results, not Glowing Spoon measurements. Benchmark percentages, F1 scores, and Elo ratings use different scales. Production evaluation also includes latency, cost, reliability, and human review effort on your workload.

Sources & methodology