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
Qwen3.8 Flash Next by alibaba is offered by 6 providers on this page. Its reference price is $0.201 per 1M input tokens and $0.50 per 1M output tokens. The lowest paid channel is Vultr at $0.10 / $0.20 per 1M, about 2.0× below the reference price.
Artificial Analysis rates it 39.8 on the Intelligence Index, with 73.1 for coding and 53.6 for agentic tasks. Against its lowest blended price of $0.125 per 1M, that is roughly 318 index points per dollar, which is the value ratio the leaderboards rank on. Median output speed is 51 tokens per second, with 2.02s to the first token. Running one task of the Artificial Analysis suite costs about $0.3722, which reflects how many tokens its reasoning consumes rather than the unit price alone.
The context window is 262,144 tokens, with an output limit of 64,000 tokens. It supports reasoning, tool use, and structured output. Accepted input modalities are Text, Image, and Video. The weights are open, so it can also be self-hosted or served through a gateway of your choice.
Quality is independently evaluated by Artificial Analysis. Speed/latency are model-level medians.
| Provider | Tier | Input | Output | Cache read | Cache write | Context | Output limit | Status |
|---|---|---|---|---|---|---|---|---|
| Vultr | Cloud | $0.10 | $0.20 | — | — | 262,144 | 131,072 | |
| AMD Qwen3.8-Flash-Next | Cloud | $0.15 | $0.47 | $0.016 | — | 262,144 | 131,072 | |
| TensorX | Gateway | $0.20 | $0.50 | $0.05 | — | 262,144 | 64,000 | |
| Requesty | Gateway | $0.20 | $0.50 | $0.05 | — | 262,144 | 262,144 | |
| Requesty qwen3.8-flash-next@eu | Gateway | $0.20 | $0.50 | $0.05 | — | 262,144 | 262,144 | |
| Cortecs | Gateway | $0.20 | $0.50 | $0.05 | — | 262,144 | 64,000 |
Sorted by blended price (input×0.75 + output×0.25) asc. The official channel always shows regardless of rank. Whether a gateway's low price is actually usable can't be verified.
| Name | Conditions | Score | Metric | Source |
|---|---|---|---|---|
| DeepSWE | harness: mini-swe-agent · v1.1 | 58.7 | score | Source ↗ |
| SWE-Bench Pro | harness: Claude Code · dataset: Qwen refined and corrected task set | 62.5 | score | Source ↗ |
| SWE-Bench Multilingual | harness: mini-swe-agent | 81 | score | Source ↗ |
| NL2Repo | harness: Claude Code | 48.1 | score | Source ↗ |
| CoWorkBench | dataset: internal | 73.9 | score | Source ↗ |
| JobBench | — | 55.7 | score | Source ↗ |
| Agents' Last Exam | — | 24.3 | pass@1 | Source ↗ |
| Agents' Last Exam | — | 51.2 | score | Source ↗ |
| Toolathlon-Verified | — | 73.5 | pass@1 | Source ↗ |
| IFBench | — | 81.3 | score | Source ↗ |
| GPQA Diamond | — | 91.7 | score | Source ↗ |
| Humanity's Last Exam | variant: without tools; GPT-4o judge | 35.9 | score | Source ↗ |
| LiveCodeBench | v6 | 91.9 | score | Source ↗ |
| ClawEval-MM | — | 64.4 | pass@3 | Source ↗ |
| ClawEval-MM | — | 60.4 | average score | Source ↗ |
| RecreationBench | dataset: internal | 49.9 | score | Source ↗ |
| AndroidWorld | — | 84.5 | score | Source ↗ |
| OSWorld | v2.0 | 19.4 | binary completion rate | Source ↗ |
| OSWorld | v2.0 | 52.3 | partial score | Source ↗ |
| Vision2Web | harness: Claude Code · variant: gpt-5.4-2026-03-05 judge | 64 | score | Source ↗ |
| ERQA | — | 72.3 | score | Source ↗ |
| LVBench | — | 76.6 | score | Source ↗ |
| RealWorldQA | — | 88.5 | score | Source ↗ |
| MathVision | variant: without code interpreter · dataset: corrected annotations | 90.6 | score | Source ↗ |
| MathVision | variant: with code interpreter · dataset: corrected annotations | 95.7 | score | Source ↗ |
| CharXiv Reasoning | variant: without code interpreter | 84.6 | score | Source ↗ |
| CharXiv Reasoning | variant: with code interpreter | 90.6 | score | Source ↗ |
The same benchmark scores very differently across harness / dataset, so the qualifying conditions must be shown together.