AI Tools lists 40 model variants; the two most notable are Qwen 2.5 72B (the largest open-source model in the lineup) and DeepSeek V3 Lite (a distilled variant of DeepSeek’s latest architecture). This is a zero-registration relay focused exclusively on open-source models.
Speed & Latency: The 68 TPS Claim
The headline feature is 68 tokens per second throughput. That’s fast. For comparison, many paid relays on OpenRouter hover around 30-50 TPS for similar model sizes. The tradeoff: latency sits at 900ms. That’s the time before the first token appears. If you’re building a chatbot where the user expects an instant response, 900ms is noticeable. If you’re batch-processing or streaming, the high throughput makes up for the slow start. Key takeaways:
- 68 TPS is excellent for open-source models; expect fast generation once the stream starts.
- 900ms TTFB is average; not ideal for real-time conversational apps.
- No SLA guarantee means speed can degrade during peak hours.
Model-by-Model Breakdown
Qwen 2.5 (7B-72B)
The flagship model. The 72B variant is the most capable here, handling complex reasoning and Chinese-language tasks well. Context window is 32,768 tokens across all variants — no tier-based cuts. Stability is acceptable for a free relay, but expect occasional timeouts during high load.
DeepSeek V3 Lite
A distilled model that punches above its weight. It’s faster than Qwen 72B due to the smaller parameter count, but reasoning depth is noticeably lower. Good for summarization and code completion. Avoid for multi-step logic.
GLM-4V
Vision-capable model. Supports image inputs. Speed is slower than text-only models — expect around 40 TPS. The 32K context window applies here too. Useful for OCR or image captioning tasks, but not for production vision pipelines.
Llama 3 (8B-70B)
The 70B variant is the most reliable for English-language tasks. The 8B variant is fast (can hit 68 TPS) but quality drops significantly. Use the 70B if you need coherent outputs.
Mistral 7B
The weakest link. Low parameter count means outputs are repetitive. Only use for trivial tasks like text classification or simple extraction. Key takeaways:
- Qwen 2.5 72B is the most capable model; use for complex reasoning.
- DeepSeek V3 Lite and Llama 3 70B are good for speed-sensitive tasks.
- Mistral 7B is only useful for very basic operations.
Pricing
| Plan | Price | Models | Notes |
|---|---|---|---|
| Free | ¥0 | All 40 variants | No registration needed; rate-limited per IP |
| Pay-as-you-go | Not specified | All 40 variants | Alipay/WeChat Pay accepted; no minimum recharge listed |
| The free tier is genuinely free — no credit card, no registration. You just hit the API endpoint. The catch: no SLA, no uptime guarantee, and the 95% uptime figure means ~36 hours of downtime per month. Not production-safe. | |||
| Bottom line: The free tier is usable for prototyping and personal projects. If you need reliability, this isn’t it. |
Pros & Cons
Pros:
- Zero registration — start coding in 10 seconds
- 68 TPS throughput — fastest among free open-source relays
- 40+ model variants across 5 base architectures
- Cross-domain API calls allowed (no CORS issues for browser apps) Cons:
- 95% uptime — expect ~1.5 days of downtime per month
- Open-source models only — no GPT-4, Claude, or Gemini
- No SLA or refund policy — you get what you get
- 900ms TTFB — not suitable for real-time chat
Verdict
AI Tools is a solid choice for three specific use cases: prototyping with open-source models, batch processing where speed matters more than reliability, and learning/education. The 68 TPS throughput is genuinely impressive for a free service. Do not use this for production. The 95% uptime and lack of SLA make it unreliable for customer-facing applications. The open-source-only model selection also limits its utility — you can’t access GPT-4 or Claude here. For developers in China who need access to open-source LLMs without a VPN and without paying, this is the fastest option available. For anything else, look at paid relays with SLAs.
FAQ
Is the free tier rate-limited?
Yes. The platform doesn’t specify exact limits, but expect throttling after ~100 requests per hour per IP. This is typical for free relays.
Can I use this with LangChain or LlamaIndex?
Yes. The API endpoint supports standard OpenAI-compatible format. Cross-domain calls are explicitly allowed, so browser-based tools work too.
Which model has the largest context window?
All models share the same 32,768 token limit. No tier-based context window cuts. This is unusual — most relays reduce context for smaller models.
How does 95% uptime affect daily usage?
Expect ~36 hours of downtime per month. The platform does not publish maintenance windows. If your app needs 99.9% uptime, this is not the right choice.
Can I switch to a paid plan for better reliability?
The platform does not currently offer paid plans with SLAs. Payment methods (Alipay/WeChat Pay) are listed, but no pricing for pay-as-you-go has been published. Check back in 2026 for updates.
Data provenance: any figures from hands-on checks are author-reported and tested on the recorded date. Independent verification is unavailable unless a source is linked.
