Cloudflare 悄悄把 GLM-5.2、Kimi K3、DeepSeek V4 都接上了，而且只要一个 token
Cloudflare Quietly Added GLM-5.2, Kimi K3, and DeepSeek V4 — All Behind One Token
一句话结论：你现在可以在 Cloudflare Worker 里，用同一行代码调用 GLM-5.2、Kimi K2.7-Code、DeepSeek V4 Pro、MiniMax M3——不用去各家申请 API key，不用管各家计费，推理还跑在离用户最近的边缘节点上。
起因
我们在做的应用要嵌 AI 能力。常规做法是：选一家模型厂商 → 申请 key → 管额度 → 写适配层 → 想换模型时再来一遍。如果同时想用几家的开源模型，这套流程就得复制几份。
所以有个很自然的问题：Cloudflare 的 Workers AI，到底能不能直接给到最新的开源模型？ 如果能，应用和 AI 就是同一个运行时里的事，而不是「应用去调远端 API」。
于是把官方文档、blog、changelog 翻了一遍。结果比预期好。
结论：你想要的基本都有
| 模型 | 有没有 | 调用 ID |
|---|---|---|
| GLM-5.2（智谱，agentic coding） | ✅ 2026-06-16 上线 | glm-5.2 |
| Kimi K2.7 Code（月之暗面，1T MoE） | ✅ 2026-06-12 上线 | kimi-k2.7-code |
| Kimi K3（1M 上下文） | ✅ | moonshotai/kimi-k3 |
| DeepSeek V4 Pro | ✅ | deepseek/deepseek-v4-pro |
| MiniMax M3 | ✅ | minimax/m3 |
| GLM 5.3 / DeepSeek V4 Flash | ❌ 暂无 | — |
除此之外还有 gpt-oss-120b/20b（OpenAI 开源权重）、llama-4-scout、qwen3-30b、gemma-4-26b、nemotron-3-120b……
调用长这样，自托管和第三方完全一致：
const res = await env.AI.run('glm-5.2', {
messages: [{ role: 'user', content: '...' }]
})
// 换成 DeepSeek V4 Pro？改一个字符串而已
const res2 = await env.AI.run('deepseek/deepseek-v4-pro', {
messages: [{ role: 'user', content: '...' }]
})不需要 DeepSeek 的 key，也不需要 Fireworks 的 key——用你自己的 Cloudflare token，Cloudflare 统一计费。这是整件事里最省心的部分。
但有一条分界线，不讲清楚会踩坑
模型分两类，调用方式一样，商业属性完全不同：
| Cloudflare 自托管 | 第三方托管 | |
|---|---|---|
| 谁的 GPU | Cloudflare 自己的 | 合作方（DeepSeek V4 Pro 走 Fireworks） |
| 价格 | 文档公开单价 | 只在 dashboard 能看到 |
| 数据 | 不出 Cloudflare | 出网到合作方 |
GLM、Kimi K2.x、Qwen、gpt-oss、Llama 4 属于自托管；DeepSeek V4 Pro、Kimi K3、MiniMax M3 属于第三方。
所以：第三方那批，决策前必须登 dashboard 查价，别拿自托管的价格去估算。涉及敏感数据的场景，也建议只用自托管那批。
价格：便宜档真的很便宜
免费额度每天 10,000 Neurons，超出后 $0.011/1000 Neurons。自托管部分单价公开：
| 模型 | 输入 $/M | 输出 $/M |
|---|---|---|
qwen3-30b-a3b-fp8 | 0.051 | 0.335 |
glm-4.7-flash | 0.060 | 0.400 |
gpt-oss-20b | 0.200 | 0.300 |
llama-4-scout-17b | 0.270 | 0.850 |
gpt-oss-120b | 0.350 | 0.750 |
kimi-k2.7-code | 0.950 | 4.000 |
glm-5.2 | 1.400 | 4.400 |
glm-4.7-flash 值得单独说：131K 上下文、$0.06/M 输入、工具调用支持 100+ 语言，便宜到可以当默认档，复杂请求再往上升级。
⚠️ glm-5.2 和 kimi-k2.* 需要 Workers Paid 计划，免费额度只够验证。
为什么这件事值得关注？
不是「又多了一个 API 聚合器」。 区别在三处：
① 推理和应用在同一个边缘节点。 不是 Worker 去调远端 API，而是模型就跑在你代码执行的地方。流式输出、多轮工具调用这类场景，首字延迟差别很直观。
② 一套凭证、一套账单。 不用为 GLM/Kimi/MiniMax/DeepSeek 各开账号、各管 key 和额度。
③ 和其余原语天然同构。 Durable Objects 存会话、KV/R2 存产物、Queues 异步、Workflows 编排、Vectorize 做 RAG、AI Gateway 做缓存限流——同一个账号同一个运行时，不用在几朵云之间搬数据。
他们是真在自己跑，不是贴牌
Cloudflare 有篇工程博客讲了怎么把这些大模型塞进自己的机器：
- KV Cache 从 BF16 量化到 FP8：缓存减半，Kimi K2.6 的上下文容量从 686K 涨到 1.37M tokens，峰值并发吞吐 +41%
- GLM 权重从 INT8 压到 INT4：体积 705GB → 421GB（-40%），同样硬件能放 1.18M tokens 的 KV cache；低并发下解码吞吐 +55%（60 → 92 tokens/s）
- prefill / decode 池分离：解码用 INT4、预填用 FP8，各取所长
- 缓存完整性校验：防止请求读到别人的 cache page，开销 <1%
这些优化说明大模型是真自托管并做了针对性工程，不是简单转发请求。
上新快，但弃用也快
看 2026 年的 changelog，节奏是每月 6–20 个模型变更：
- 6-16 GLM-5.2 上线
- 6-12 Kimi K2.7 Code 上线（四天内两个大模型）
- 5-08 一次性弃用 19 个老模型（Llama 3/3.1、Mistral 7B、Gemma 7B、Phi-2…）
- 4-20 Kimi K2.6 · 4-04 Gemma 4 · 3-19 Kimi K2.5 · 3-11 Nemotron 3 · 2-13 GLM-4.7-Flash
开源模型发布后通常数周内就能用上——这是好消息。但 5 月那次一口气砍掉 19 个模型也提醒我们：生产代码不要硬编码模型 ID，留一层映射，并盯住 changelog 的 deprecation 公告。
怎么开始？
- 先用便宜档验证形态：
qwen3-30b-a3b-fp8或glm-4.7-flash，免费额度就能跑通端到端 - 需要强推理/长上下文再升级到
glm-5.2或kimi-k2.7-code - 1M 上下文场景（整仓代码、长文档）才考虑
moonshotai/kimi-k3——先查价 - 架构上留模型映射层，配合 AI Gateway 做缓存、限流、可观测
常见问题
Q：不用自己申请 DeepSeek/Kimi 的 key，是不是意味着完全免费？ 不是。第三方托管的模型（DeepSeek V4 Pro、Kimi K3、MiniMax M3）仍按量计费，只是账单统一走 Cloudflare，具体单价要登 dashboard 查，文档不公开。
Q：敏感数据能用第三方托管的模型吗？ 不建议。第三方托管意味着推理请求会出网到合作方基础设施（如 Fireworks），涉及敏感数据的场景应只用 Cloudflare 自托管的那批模型。
Q：生产环境该怎么防止模型被弃用影响服务？ 不要在业务代码里硬编码模型 ID，做一层模型映射配置，并关注 Workers AI 的 changelog——2026-05-08 曾一次性弃用 19 个老模型。
参考资料
- Workers AI 模型目录：https://developers.cloudflare.com/workers-ai/models/
- Cloudflare AI 全模型目录（214 个）：https://developers.cloudflare.com/ai/models/
- Workers AI 定价：https://developers.cloudflare.com/workers-ai/platform/pricing/
- Workers AI Changelog：https://developers.cloudflare.com/workers-ai/changelog/
- Cloudflare 工程博客《Smaller, faster, safer — running Kimi and GLM at scale》：https://blog.cloudflare.com/smaller-faster-safer-models/
- Cloudflare 博客《Powering the agents — Workers AI now runs large models》：https://blog.cloudflare.com/workers-ai-large-models/
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。
TL;DR: You can now call GLM-5.2, Kimi K2.7-Code, DeepSeek V4 Pro, and MiniMax M3 from a Cloudflare Worker using the same line of code — no need to apply for API keys from each vendor, no need to juggle separate billing, and inference runs on the edge node closest to your user.
Why We Looked Into This
Our app needs embedded AI capability. The usual playbook is: pick a model vendor → apply for a key → manage quotas → write an adapter layer → repeat when you want to switch models. Want to use several vendors’ open models at once? Multiply that process.
So a natural question came up: can Cloudflare’s Workers AI directly serve the latest open-source models? If it can, “app” and “AI” become one runtime instead of “app calling a remote API.”
We went through the official docs, blog posts, and changelog. The result was better than expected.
The Answer: Basically Everything You Want Is There
| Model | Available? | Model ID |
|---|---|---|
| GLM-5.2 (Zhipu, agentic coding) | ✅ launched 2026-06-16 | glm-5.2 |
| Kimi K2.7 Code (Moonshot AI, 1T MoE) | ✅ launched 2026-06-12 | kimi-k2.7-code |
| Kimi K3 (1M context) | ✅ | moonshotai/kimi-k3 |
| DeepSeek V4 Pro | ✅ | deepseek/deepseek-v4-pro |
| MiniMax M3 | ✅ | minimax/m3 |
| GLM 5.3 / DeepSeek V4 Flash | ❌ not yet | — |
Also available: gpt-oss-120b/20b (OpenAI’s open weights), llama-4-scout, qwen3-30b, gemma-4-26b, nemotron-3-120b, and more.
The call shape is identical for self-hosted and third-party models:
const res = await env.AI.run('glm-5.2', {
messages: [{ role: 'user', content: '...' }]
})
// Switch to DeepSeek V4 Pro? Just change one string.
const res2 = await env.AI.run('deepseek/deepseek-v4-pro', {
messages: [{ role: 'user', content: '...' }]
})No DeepSeek key, no Fireworks key needed — you use your own Cloudflare token, and Cloudflare handles billing centrally. This is the most convenient part of the whole thing.
One Dividing Line You Need to Know, Or You’ll Get Burned
The models split into two categories that call the same way but have completely different commercial properties:
| Cloudflare self-hosted | Third-party hosted | |
|---|---|---|
| Whose GPU | Cloudflare’s own | Partner infrastructure (DeepSeek V4 Pro runs on Fireworks) |
| Price | Publicly documented | Only visible in the dashboard |
| Data | Never leaves Cloudflare | Egresses to the partner |
GLM, Kimi K2.x, Qwen, gpt-oss, and Llama 4 are self-hosted; DeepSeek V4 Pro, Kimi K3, and MiniMax M3 are third-party.
So: for the third-party batch, you must log into the dashboard and check pricing before deciding — don’t estimate cost using self-hosted pricing. For sensitive-data workloads, stick to the self-hosted batch.
Pricing: The Cheap Tier Is Genuinely Cheap
Free tier is 10,000 Neurons/day; beyond that it’s $0.011 per 1,000 Neurons. Self-hosted pricing is public:
| Model | Input $/M | Output $/M |
|---|---|---|
qwen3-30b-a3b-fp8 | 0.051 | 0.335 |
glm-4.7-flash | 0.060 | 0.400 |
gpt-oss-20b | 0.200 | 0.300 |
llama-4-scout-17b | 0.270 | 0.850 |
gpt-oss-120b | 0.350 | 0.750 |
kimi-k2.7-code | 0.950 | 4.000 |
glm-5.2 | 1.400 | 4.400 |
glm-4.7-flash deserves a callout: 131K context, $0.06/M input, tool calling in 100+ languages — cheap enough to be your default tier, upgrading only for complex requests.
⚠️ glm-5.2 and kimi-k2.* require a Workers Paid plan — the free tier is only enough for validation.
Why Does This Matter?
This isn’t “yet another API aggregator.” The difference is in three places:
① Inference and app share the same edge node. The Worker isn’t calling a remote API — the model runs right where your code executes. For streaming output and multi-turn tool calling, the first-token latency difference is very noticeable.
② One credential, one bill. No separate accounts, keys, or quotas for GLM/Kimi/MiniMax/DeepSeek.
③ Native composability with the rest of the primitives. Durable Objects for sessions, KV/R2 for artifacts, Queues for async work, Workflows for orchestration, Vectorize for RAG, AI Gateway for caching and rate limiting — all in the same account, same runtime, no shuffling data between clouds.
They’re Actually Running It Themselves, Not Just Reselling It
Cloudflare published an engineering blog post on how they fit these large models onto their own hardware:
- KV cache quantized from BF16 to FP8: cache halved, Kimi K2.6’s context capacity grew from 686K to 1.37M tokens, peak concurrent throughput +41%
- GLM weights compressed from INT8 to INT4: size dropped from 705GB to 421GB (-40%), the same hardware now fits 1.18M tokens of KV cache; low-concurrency decode throughput +55% (60 → 92 tokens/s)
- Separate prefill/decode pools: INT4 for decode, FP8 for prefill, each optimized for its job
- Cache integrity verification: prevents a request from reading another tenant’s cache page, at <1% overhead
These optimizations show this is genuine self-hosting with targeted engineering, not simple request forwarding.
New Models Arrive Fast, But Deprecations Are Fast Too
Looking at the 2026 changelog, the pace is 6–20 model changes per month:
- 6-16 GLM-5.2 launched
- 6-12 Kimi K2.7 Code launched (two major models in four days)
- 5-08 19 old models deprecated at once (Llama 3/3.1, Mistral 7B, Gemma 7B, Phi-2…)
- 4-20 Kimi K2.6 · 4-04 Gemma 4 · 3-19 Kimi K2.5 · 3-11 Nemotron 3 · 2-13 GLM-4.7-Flash
Open-source models typically become available within weeks of release — that’s the good news. But the May deprecation of 19 models in one shot is a reminder: don’t hardcode model IDs in production code. Keep a mapping layer, and watch the changelog’s deprecation notices.
How to Get Started
- Validate the shape with the cheap tier first:
qwen3-30b-a3b-fp8orglm-4.7-flash— the free tier is enough for an end-to-end test - Upgrade to
glm-5.2orkimi-k2.7-codewhen you need stronger reasoning or longer context - Only consider
moonshotai/kimi-k3for 1M-context use cases (whole-repo code, long documents) — check pricing first - Keep a model-mapping layer in your architecture, paired with AI Gateway for caching, rate limiting, and observability
FAQ
Q: Not needing separate DeepSeek/Kimi keys — does that mean it’s free? No. Third-party-hosted models (DeepSeek V4 Pro, Kimi K3, MiniMax M3) are still billed by usage; billing is just consolidated through Cloudflare. Exact pricing isn’t published and must be checked in the dashboard.
Q: Can I use the third-party-hosted models for sensitive data? Not recommended. Third-party hosting means inference requests egress to partner infrastructure (e.g., Fireworks). Sensitive-data workloads should stick to Cloudflare’s self-hosted models.
Q: How do I protect production from model deprecations? Don’t hardcode model IDs in business logic — keep a model-mapping config layer and watch the Workers AI changelog. On 2026-05-08, Cloudflare deprecated 19 old models in a single announcement.
References
- Workers AI model catalog: https://developers.cloudflare.com/workers-ai/models/
- Cloudflare AI full model catalog (214 models): https://developers.cloudflare.com/ai/models/
- Workers AI pricing: https://developers.cloudflare.com/workers-ai/platform/pricing/
- Workers AI changelog: https://developers.cloudflare.com/workers-ai/changelog/
- Cloudflare engineering blog, “Smaller, faster, safer — running Kimi and GLM at scale”: https://blog.cloudflare.com/smaller-faster-safer-models/
- Cloudflare blog, “Powering the agents — Workers AI now runs large models”: https://blog.cloudflare.com/workers-ai-large-models/
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18，不接受商业合作、不代表任何企业或机构立场，也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息，而是给出有独立思考的深入分析，希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型，仅供学习交流与技术借鉴。它们大多仍处于早期阶段，有待进一步研究和验证，请勿直接用于工作或生产环境；如需采用，请先自行充分测试，并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理，不代表文中提及的任何公司、产品、模型的官方立场，未与其构成商业关联或合作关系。
- 科技行业信息更新极快，我们尽力保证内容准确、及时，但不对完整性、实时性做绝对保证，具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有，我们会尽量注明来源；如你认为存在版权疑问或侵权，请通过下方邮箱联系我们，收到通知后会尽快核实处理（更正、加注来源或删除）。
- 文章内容仅为技术科普与个人观点，不构成投资、法律或其他专业建议，据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询：[email protected]
💬 评论与讨论
使用 GitHub 账号登录后发表评论