API model string
not published (delisted from current DashScope catalog)Context window
128K tokens (Qwen2-72B)
Max output not published
- Knowledge cutoff
- not published
- Released
- 2024
- Modalities
- text in, text out (Qwen2-VL adds vision)
Last verified July 10, 2026
Rate limits by tier
Qwen2 is delisted from the current DashScope international pricing catalog (July 1-2, 2026 update). No first-party API endpoint with published limits exists. The table below describes the real access paths: third-party hosting and self-hosted open weights.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| DashScope first-party — NOT AVAILABLE | n/a | not published (delisted) | not published | not published | not published | Qwen2 is no longer in the current DashScope international pricing catalog. First-party price is not published; endpoint may return errors. Alibaba's 1M free trial does not apply to delisted models. |
| Third-party hosts (DeepInfra, Fireworks, Together AI) | Account on the relevant third-party inference host | varies by host (not published here) | varies by host | varies by host | varies by host | Qwen2.5-72B (the most recent Qwen2-family model) is still widely served on third-party hosts. Rates are historically 50-80% below former DashScope prices (per third-party trackers — verify current rates on each host before quoting). |
| Self-host (open weights) | Own GPU infrastructure; Apache 2.0 license (commercial use permitted) | self-managed | self-managed | self-managed | self-managed | Qwen2-72B-Instruct open weights available at huggingface.co/Qwen/Qwen2-72B-Instruct. Apache 2.0 license permits commercial use. Hardware requirement: approximately 2-4 A100 40GB GPUs at FP16. |
Swipe the table sideways to see every limit column.
- 1.Qwen2.5 (72B, Coder, VL variants) is still widely served on third-party hosts; Qwen2 proper is largely superseded by Qwen2.5 and now Qwen3.
- 2.Open weights on Hugging Face: huggingface.co/Qwen/Qwen2-72B-Instruct (Apache 2.0 license).
- 3.Alibaba's new 1M-token free trial (for current-catalog models) does not apply to Qwen2 — the model is delisted.
- 4.No first-party Qwen2 API price is published — the model is delisted from DashScope international as of July 2026.
- 5.Third-party indicative pricing for Qwen2.5-72B (verify on each host before quoting): DeepInfra historically serves open-weight models at a significant discount vs prior DashScope rates. Exact current Qwen2 per-token rate is not confirmed.
- 6.For migration planning, see successor Qwen3-Max on DashScope international: $1.20 input / $6.00 output per MTok; 50% batch discount (cannot combine with context caching).
Limits verified against the Alibaba (DashScope) docs, July 10, 2026.
not published (delisted from current DashScope catalog) vs the alternatives
Qwen2 (delisted) compared to its own successor and live alternatives for teams evaluating migration options.
| Aspect | not published (delisted from current DashScope catalog) | Qwen3-Max | Llama 4 Scout | Gemma 3 |
|---|---|---|---|---|
| First-party API availability | delisted (no first-party) | live DashScope | third-party hosts (DeepInfra, Groq) | Google AI Studio / Vertex |
| Context window | 128K (Qwen2-72B) | 262K (Qwen3-Max) | up to 1M natively (Llama 4 Scout) | 131K (Gemma 3) |
| License | Apache 2.0 (open weights, commercial use OK) | Qwen3 License | Llama 4 Community License | Gemma License |
| Input $/MTok | not published (delisted) | $1.20 (DashScope) | $0.08-$0.11 (DeepInfra/Groq) | $0.040 (Google) |
| Output $/MTok | not published (delisted) | $6.00 | $0.30 | $0.080 |
| Self-host option | yes (Apache 2.0, open weights) | yes (Qwen3 open weights) | yes (Meta open weights) | yes (Google open weights) |
| Multimodal | Qwen2-VL yes; base Qwen2 text-only | text-only (Qwen3-Max base) | text + image (Llama 4 Scout) | text + image (Gemma 3) |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
No first-party Qwen2 API endpoint exists — 429 handling is per third-party host.HTTP 429 Too Many Requests (standard on DeepInfra, Fireworks, Together AI with Retry-After header)ThrottlingException or RateLimitExceeded (host-specific; check your provider's docs)Why it happens & how to fix it
Using a decommissioned DashScope Qwen2 endpoint
Switch to Qwen3-Max on DashScope (model string: qwen3-max) or Qwen2.5-72B on a current third-party host such as DeepInfra or Fireworks.
Third-party host RPM limit exceeded
Implement exponential backoff with jitter; check the host's rate-limit response headers (Retry-After or x-ratelimit-reset) and honor them.
Account balance depleted on third-party host
Add credits to your third-party host account (DeepInfra, Fireworks, or Together AI billing portal).
Retry strategy
Standard HTTP 429 handling applies on all third-party hosts (DeepInfra, Fireworks, Together). Honor the Retry-After header if present. Otherwise, use exponential backoff: wait = min(60s, 2^attempt + random(0, 1)s). Most third-party hosts use OpenAI-compatible APIs — the same retry wrapper works across hosts by changing base_url.
1import OpenAI from 'openai';23// Swap base_url for your host:4// DeepInfra: https://api.deepinfra.com/v1/openai (model: Qwen/Qwen2-72B-Instruct)5// Fireworks: https://api.fireworks.ai/inference/v1 (model: accounts/fireworks/models/qwen2-72b-instruct)6// For new projects, consider migrating to Qwen3-Max on DashScope.7const client = new OpenAI({8 apiKey: process.env.HOST_API_KEY ?? '',9 baseURL: 'https://api.deepinfra.com/v1/openai',10});1112async function callWithRetry(13 prompt: string,14 maxAttempts = 515): Promise<string> {16 for (let attempt = 0; attempt < maxAttempts; attempt++) {17 try {18 const response = await client.chat.completions.create({19 model: 'Qwen/Qwen2-72B-Instruct',20 messages: [{ role: 'user', content: prompt }],21 });22 return response.choices[0].message.content ?? '';23 } catch (err: any) {24 if (err?.status === 429 && attempt < maxAttempts - 1) {25 const retryAfter = err.headers?.['retry-after'];26 const wait = retryAfter27 ? parseInt(retryAfter, 10) * 100028 : Math.min(60000, 1000 * Math.pow(2, attempt)) + Math.random() * 1000;29 console.warn(`429 on attempt ${attempt + 1}; retrying in ${wait}ms`);30 await new Promise((r) => setTimeout(r, wait));31 } else {32 throw err;33 }34 }35 }36 throw new Error('Max retry attempts reached');37}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Third-party free tier / trial credits
ImmediateSign up on DeepInfra (deepinfra.com) or Fireworks (fireworks.ai) — both offer trial credits for new accounts
Unlocks: Limited free inference on Qwen2.5-72B (most recent Qwen2-family model still widely hosted)
Third-party pay-as-you-go
ImmediateAdd payment method on DeepInfra, Fireworks, or Together AI
Unlocks: Per-token access at host rates (verify current Qwen2/Qwen2.5 pricing on each host — not published here as rates change frequently)
Self-host (open weights)
GPU infrastructure setup timeDownload Qwen/Qwen2-72B-Instruct from huggingface.co/Qwen/Qwen2-72B-Instruct; Apache 2.0 license, commercial use permitted
Unlocks: No per-token cost; full control; approximately 2-4 A100 40GB GPUs required at FP16
Migrate to Qwen3-Max (recommended)
Immediate (API switch only)Update model string to qwen3-max on DashScope international; add USD balance
Unlocks: Active DashScope catalog support, 262K context, batch 50% off, thinking mode, ongoing pricing transparency
Cut your token spend
Migrate to Qwen3-Max or Qwen3.5-Plus for first-party DashScope support
Eliminates uncertainty; access to published pricing and ongoing model supportQwen2 is delisted from the current DashScope catalog. Switch to qwen3-max (262K context, $1.20/$6.00, batch 50% off) or qwen3.5-plus for first-party pricing transparency and active support.
Use Qwen2.5-72B on budget-friendly third-party hosts for cost-sensitive workloads
Historically 50-80% below prior DashScope rates per third-party trackersDeepInfra and Fireworks serve open-weight Qwen2.5-72B at discounted rates (verify current pricing on each host before quoting — rates change and are not confirmed here as of July 2026).
Self-host for very high token volumes
Zero per-token cost; GPU infrastructure replaces API spendApache 2.0 license permits commercial self-hosting. At ~50M+ tokens/month, amortized GPU cost may beat third-party per-token rates. Deploy with vLLM for throughput efficiency.
Use Qwen2-VL variant if multimodal capability is needed
Avoids switching providers for vision tasksQwen2-VL (vision-language) is still available on some third-party hosts and as open weights. If vision is a hard requirement and budget is constrained, evaluate Qwen2-VL before migrating entirely to Qwen3-Max.
Pin model version string on third-party hosts
Prevents silent drift to a different checkpointWhen using third-party inference hosts, always specify the full model identifier (e.g., Qwen/Qwen2-72B-Instruct) rather than a generic alias, to avoid silent updates to a different checkpoint.
Frequently asked questions
Does Qwen2 still have an API?
Not through DashScope first-party as of July 2026. Qwen2 was delisted from the DashScope international catalog during the July 1-2, 2026 pricing update. No endpoint with a published price exists. Access is available through third-party hosts (DeepInfra, Fireworks, Together AI) for Qwen2.5-72B, or via self-hosted open weights (Apache 2.0 license).
Is Qwen2 API free?
No free hosted API exists for Qwen2 — it is delisted from DashScope. Alibaba's 1M-token free trial applies only to models in the current DashScope catalog. You can run Qwen2-72B for free by self-hosting the open weights (Apache 2.0 license), paying only for GPU infrastructure.
What is the best migration path from Qwen2?
For most teams: switch to Qwen3-Max on DashScope international (model string: qwen3-max). It is the current DashScope flagship, offers 262K context (vs 128K for Qwen2-72B), costs $1.20/$6.00 per MTok, and supports a 50% batch discount. If your workload is cost-sensitive and context requirements are modest, also evaluate Llama 4 Scout at $0.08-$0.11/$0.30 on third-party hosts.
Can I still self-host Qwen2?
Yes — Qwen2-72B-Instruct is available on Hugging Face at huggingface.co/Qwen/Qwen2-72B-Instruct under the Apache 2.0 license, which permits commercial use. You need approximately 2-4 A100 40GB GPUs at FP16. Deploy with vLLM for efficient throughput.
What is Qwen2 vs Qwen2.5 vs Qwen3-Max?
Qwen2 was released in 2024 and is now largely superseded. Qwen2.5 added capability improvements and is still served by third-party hosts. Qwen3-Max (released January 2026) is the current prior-gen DashScope flagship with 262K context and thinking mode, now itself superseded by Qwen3.7-Max. For new projects, use Qwen3-Max or Qwen3.7-Max on DashScope.
Why is Qwen2 rate limit information not available?
Because Qwen2 is delisted from the current DashScope catalog, no rate limits are published. DashScope did not prominently publish RPM/TPM limits for Qwen2 even when it was active. Third-party hosts that serve Qwen2.5-72B have their own rate limits visible in their consoles.
How do I integrate Qwen2 for existing workflows that I cannot immediately migrate?
Use a third-party OpenAI-compatible host: DeepInfra (base_url: https://api.deepinfra.com/v1/openai, model: Qwen/Qwen2-72B-Instruct) or Fireworks (base_url: https://api.fireworks.ai/inference/v1, model: accounts/fireworks/models/qwen2-72b-instruct). This approach buys migration time. RapidDev (rapidevelopers.com/contact) can help architect a phased migration to Qwen3-Max while maintaining backward compatibility for existing integrations.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.