一个说法正在流传：NVIDIA 免费发放 Kimi K3、DeepSeek V4.1 Flash 和两个版本 GLM 5.3 的 API key——无需信用卡，十五分钟搞定。作者亲自走了一遍流程，结论是：模型是真的，免费额度比转发说的还慷慨，但标题里漏掉了一个关键点——手机号验证，而且并非所有国家都支持。
NVIDIA 送什么
NVIDIA Build 是 NVIDIA 托管的开放模型目录。API 兼容 OpenAI 格式：同样是 /v1/chat/completions 请求，只需换 base URL 和 key。截至 2026 年 10 月 1 日，integrate.api.nvidia.com/v1/models 列出了 81 个模型，包括今秋最受关注的四款开源权重发布。
大多数转发搞错的地方是额度。新账户过去大约获得 1000 积分，现在积分制已取消。NVIDIA 账户验证对话框现在显示"API 请求无每日限制"。真正受限的是速率：NVIDIA 开发者论坛上的讨论普遍认为每个 key 每分钟约 40 次请求，论坛版主 7 月表示限制取决于模型、使用场景和整体流量，免费层无法官方提高。
40 RPM 意味着如果不停调用，一天就是 57,600 次请求。对于原型开发、个人智能体或在自己的任务上评估模型来说，这实际上等于无限。如果要用于生产，NVIDIA 会引导你使用付费方案：合作伙伴端点或自行部署。
四款模型
这四款都是混合专家（MoE）模型，且都拥有 1,048,576 token 的上下文窗口。Kimi K3（moonshotai/kimi-k3）约 2.8T 总参数量，主打长程智能体编码、工具调用、图像输入，思维链常开；GLM 5.3（z-ai/glm-5.3）753B 参数，擅长文本、推理、工具调用；GLM 5.3 Flash（z-ai/glm-5.3-flash）320B 参数，支持文本+图像、工具、结构化输出；DeepSeek V4.1 Flash（deepseek-ai/deepseek-v4.1-flash）552B 参数，多模态，推理强度可调 1-100。四款均采用 MIT 许可。
作者建议：写代码首选 Kimi K3——它是这波热度的原因，而且自己本地跑需要一整台机架，免费使用别人 GPU 上的算力是大多数人的唯一合理方式。GLM 5.3 Flash 和 DeepSeek V4.1 Flash 则适合快速、便宜的调用，不需要最聪明的答案。
隐藏条件：手机验证
注册流程本身很顺畅：邮箱、密码、NVIDIA 开发者计划账户，按承诺无需银行卡。但在 API key 页面会弹出一个窗口——"我们需要验证您的手机号码"——必须输入一次性短信验证码才能拿到 key。NVIDIA 称这是防欺诈与滥用保护。
国家列表并非全球覆盖。列表上完全没有俄罗斯；NVIDIA 在列表下方注明"正在快速扩大全球可用性"。哈萨克斯坦在列表中，但作者用哈萨克斯坦号码尝试时，"发送验证码"按钮始终没有激活，页面也没有报错。作者没有绕过验证，也不建议这样做：虚拟号码和借用 SIM 卡违反条款，用这种方式获得的 key 可能连账号一起被封。
所以"十五分钟、无卡"只有在你的手机号来自支持国家时才成立。
首次请求
如果你的号码受支持，第一次调用只需一分钟。任何允许修改 base URL 的客户端都可以用，包括编码智能体。只需将模型 ID 换成表中的任意一个即可。Kimi K3 有一个细节：思维链始终开启，在多轮对话或工具调用中，必须把上一条完整的助手消息（包括 reasoning_content 和 tool_calls）发回去，否则它会丢失上下文。
结论
这个优惠本身很棒：本季最强的四个开放模型，无需付费、没有每日上限、标准 API。如果你的手机号符合条件，这是无需租硬件就能体验 Kimi K3 的最简单方式。但要知道，"无卡、十五分钟"悄悄跳过了手机这一步——而这一步正是让一些人止步的地方。
本文来源：https://dev.to/klukyanov/kimi-k3-deepseek-and-glm-free-from-nvidia-i-tested-the-claim-and-here-is-the-catch-2m9b
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
A rumor is circulating: NVIDIA is giving away free API keys for Kimi K3, DeepSeek V4.1 Flash, and two versions of GLM 5.3 — no credit card required, ready in 15 minutes. I walked through the process and can confirm: the models are real, the free quota is even more generous than advertised, but the headline misses a crucial detail — phone verification is required, and not all countries are supported.
What NVIDIA is giving away
NVIDIA Build is NVIDIA’s hosted repository of open models. The API is compatible with OpenAI’s format: same `/v1/chat/completions` request, just change the base URL and key. As of October 1, 2026, `integrate.api.nvidia.com/v1/models` lists 81 models, including four highly anticipated open-weight releases from this fall.
Most posts misunderstand the quota. New accounts used to get ~1,000 credits; now, the credit system is gone. The NVIDIA account verification dialog now explicitly states “Zero Daily Limit on API Requests.” The only limit is rate: 40 requests per minute per key (according to NVIDIA forums), which a forum moderator confirmed in July “depends on model, use case, and overall traffic” — free tiers can’t be officially increased.
40 RPM means 57,600 requests per day if you run nonstop — effectively unlimited for prototypes, personal agents, or evaluating models on your data. For production, NVIDIA guides you toward paid endpoints or self-hosting.
The four models
These are all Mixture of Experts (MoE) models with 1,048,576 token context windows. Kimi K3 (moonshotai/kimi-k3) has ~2.8T total parameters, excels at long-range agentic coding, tool use, and image input; GLM 5.3 (z-ai/glm-5.3) at 753B parameters, strong in text, reasoning, and tools; GLM 5.3 Flash (z-ai/glm-5.3-flash) at 320B, supports text+image, tools, and structured output; DeepSeek V4.1 Flash (deepseek-ai/deepseek-v4.1-flash) at 552B, multimodal with tunable reasoning intensity (1–100). All are MIT-licensed.
My recommendation: for coding, pick Kimi K3 — it’s the reason this hype exists, and local running requires a full rack; free access to someone else’s GPUs makes sense. GLM 5.3 Flash and DeepSeek V4.1 Flash are better for fast, cheap calls when you don’t need the smartest answer.
The hidden condition: phone verification
The signup flow looks clean: email, password, NVIDIA developer account — no bank card as promised. But when you go to the API key page, a popup appears: “We need to verify your phone number.” You must enter a one-time SMS code to get the key. NVIDIA says this is for fraud and abuse prevention.
The country list is not global. Russia is entirely absent; NVIDIA notes “rapidly expanding global availability” below the list. Kazakhstan is listed, but when I tried, the “send verification code” button remained inactive, and no error message appeared — suggesting temporary non-availability or region-specific blocking. I didn’t bypass it — and you shouldn’t either: using virtual numbers or borrowed SIMs violates terms, and your key may be revoked along with your account.
So “no card, 15 minutes” is technically true only if your phone number is in the supported range.
First request: quick start
If your number is supported, the first call takes one minute: any client that lets you edit base URLs works, including coding agents. Just swap in the model ID from the table. Kimi K3 note: its chain-of-thought is always on; in multi-turn conversations or tool use, you must pass back the full assistant message (including `reasoning_content` and `tool_calls`) or it loses context.
Verdict
This is a fantastic offer: four of the most powerful open models released this quarter, with no payment, no daily caps, and standard API compatibility. If your phone number qualifies, this is the easiest way to try Kimi K3 without buying hardware. But don’t ignore the fine print: “no card, 15 minutes” glosses over the phone step — and that step may leave some users out in the cold.