Qwen (Alibaba)
Listed at https://qwen.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Qwen3.8-Max: ¥8/M, Qwen3.5-Max: ¥4/M, Qwen3.5-Plus: ¥2/M | per million tokens |
| Output | Qwen3.8-Max: ¥24/M, Qwen3.5-Max: ¥12/M, Qwen3.5-Plus: ¥6/M | per million tokens |
🤖 Supported Models (16)
✨ Pros
- ✓Qwen3.8-Flash (hosted, 2026-08-26): $0.15/$0.47 per 1M tokens, multimodal, 1M context, OpenAI-compatible
- ✓Qwen3.8-Flash-Next (open weights, 2026-08-26): 125B/6B MoE with Gated DeltaNet + Qwen Sparse Attention hybrid, Qwen4 architecture preview, qwen-community-1.0
- ✓Qwen3.8-Max 2.4T open-source flagship (released 2026-08-03), tops coding + multi-agent benchmarks
- ✓Best open-source ecosystem (HuggingFace leading downloads, self-hostable 7B-110B tier)
- ✓Aliyun Bailian: model serving + RAG + Agent full stack; direct China access
- ✓Qwen-Omni-Turbo multimodal (text+image+audio); Qwen-VL-Plus vision
- ✓Global routes via OpenRouter / HuggingFace / Aliyun International
⚠️ Cons
- ×Qwen3.8-Max pricing is 4× Qwen3.5-Plus (flagship positioning)
- ×Bailian platform complexity (model serving > app builder), steep learning curve
- ×Overseas users need Aliyun International (separate pricing)
- ×Native Function Calling less stable than OpenAI / Anthropic
- ×Stream mode occasional latency jitter at long context
🎯 Best For
Enterprise AI apps in China; scenarios requiring Qwen-Max top quality output; existing Alibaba Cloud customers; self-hosted open weights
💰 Pricing & Plans
| Model | Input ($/M tokens) | Output ($/M tokens) | Notes |
|---|---|---|---|
| Qwen3.8-Max | $1.10 | $3.30 | Flagship 2.4T, released 2026-08-03 |
| Qwen3.8-Max-Preview | $1.10 | $3.30 | Preview tier (same rate) |
| Qwen3.5-Max | $0.55 | $1.65 | Previous-gen flagship |
| Qwen3.5-Plus | $0.28 | $0.83 | Mid-tier, best $/perf |
| Qwen3.5-72B-Instruct | $0.55 | $1.65 | Open-weight, self-hostable |
| Qwen3.5-7B-Instruct | $0.07 | $0.21 | Open-weight, edge deployable |
| QwQ-32B | $0.28 | $0.83 | Reasoning specialist |
| Qwen-Coder-Plus | $0.28 | $0.83 | Code-tuned variant |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible /v1/chat/completions on Aliyun Bailian and OpenRouter; native DashScope SDK for advanced features (function calling, multimodal).
- •SDK Availability: Official DashScope Python/Node SDKs; community SDKs for Go, Java, Rust; OpenAI SDK works out-of-box via base_url override.
- •Tool Use: Function calling via Qwen-Agent framework; stable for single-tool workflows, multi-tool selection can require prompt engineering.
- •Context Length: Qwen3.8-Max: 128K context; Qwen3.5-Max/Plus: 128K; long-context benchmark strong (Needle-in-Haystack >99% at 100K).
- •Open Weights: Qwen3.5 series fully open (Apache 2.0 for most variants); 7B/14B/32B/72B/110B sizes; vLLM / TGI / llama.cpp deployment.
- •Streaming: SSE streaming supported; DashScope SDK handles reconnect automatically; visual reasoning via Qwen-VL-Plus streaming.
- •Batch API: Available via DashScope batch endpoint at 50% discount; 24h SLA for completion.
🧠 Coding & Agent Capabilities
Qwen3.8-Max (2.4T parameters, open-source, released 2026-08-03) topped both coding and multi-agent collaboration benchmarks on HN/Reddit buzz, becoming the first open-weight model to lead SWE-bench Polyglot and Aider polyglot evaluations. The model supports 128K context with strong needle-in-haystack (>99% at 100K), making it competitive with Claude Opus 5 / GPT-5.6 for long-context agent tasks. For agent workflows, Qwen3.8-Max integrates with the Qwen-Agent framework (open-source Python SDK) which provides built-in tool use, memory, and multi-agent orchestration. The Qwen-Coder-Plus variant is fine-tuned for code completion and refactoring. Qwen3.5-72B-Instruct remains the sweet spot for self-hosted deployments — strong coding performance at a fraction of cloud-API cost.
🌐 Regional Availability & Latency
Qwen has the strongest China-direct access of any major lab — Aliyun Bailian (bailian.console.aliyun.com) provides stable, low-latency access from mainland China with 50-150ms typical response times for short prompts. For overseas users, Aliyun International (bailian-ss.aliyun.com) offers the same models at independent pricing, plus OpenRouter and HuggingFace Inference Endpoints provide global access. From Europe/US, direct access to Aliyun International is generally 200-400ms; OpenRouter adds ~50-100ms routing overhead but offers unified billing across providers. For production workloads serving China-based users at scale, Aliyun Bailian remains the standard. Self-hosting Qwen3.5 (Apache 2.0) on a 4×A100 cluster eliminates API latency entirely for high-volume workloads.