Hosting AI models is an interesting challenge: it requires a delicate balance between software and very, very expensive hardware. At Cloudflare, we’re good at squeezing every bit of efficiency out of our hardware through clever software engineering. This is a deep dive on how we’re laying the foundation to run extra-large language models.

As we mentioned in our previous Kimi K2.5 blog post, we’re using a variety of hardware configurations in order to best serve models. A lot of hardware configurations depend on the size of inputs and outputs that users are sending to the model. For example, if you are using a model to write fanfiction, you might give it a few small prompts (input tokens) while asking it to generate pages of content (output tokens). 

Conversely, if you are running a summarization task, you might be sending in hundreds of thousands of input tokens, but only generating a small summary with a few thousand output tokens. Presented with these opposing use cases, you have to make a choice — should you tune your model configuration so it’s faster at processing input tokens, or faster at generating output tokens?

When we launched large language models on Workers AI, we knew that most of the use cases would be used for agents. With agents, you send in a large number of input tokens. It starts off with a large system prompt, all the tools, MCPs. With the first user prompt, that context keeps growing. Each new prompt from the user sends a request to the model, which consists of everything that was said before — all the previous user prompts, assistant messages, code generated, etc. For Workers AI, that means we had to focus on two things: fast input token processing and fast tool calling.

One hardware configuration that we use to improve performance and efficiency is disaggregated prefill. There are two stages to processing an LLM request: prefill, which processes the input tokens and populates the KV cache, and decode, which generates output tokens. Prefill is usually compute bound, while decode is memory bound. This means that the parts of the GPU that are used in each stage are different, and since prefill is always done before decode, the stages block one another. Ultimately, it means that we are not efficiently utilizing all of our GPU power if we do both prefill and decode on a single machine.

With prefill decode disaggregation, separate inference servers are run for each stage. First, a request is sent to the prefill stage which performs prefill and stores it in its KV cache. Then the same request is sent to the decode server, with information about how to transfer the KV cache from the prefill server and begin decoding. This has a number of advantages, because it allows the servers to be tuned independently for the role they are performing, scaled to account for more input-heavy or output-heavy traffic, or even to run on heterogeneous hardware.

This architecture requires a relatively complex load balancer to achieve. Beyond just routing the requests as described above, it must rewrite the responses (including streaming SSE) of the decode server to include information from the prefill server such as cached tokens. To complicate matters, different inference servers require different information to initiate the KV cache transfer. We extended this to implement token-aware load balancing, in which there is a pool of prefill and decode endpoints, and the load balancer estimates how many prefill or decode tokens are in-flight to each endpoint in the pool and attempts to spread this load evenly. 

After our public model launch, our input/output patterns changed drastically again. We took the time to analyze our new usage patterns and then tuned our configuration to fit our customer’s use cases.

Here’s a graph of our p90 Time to First Token drop after shifting traffic to our new PD disaggregated architecture, whilst request volume increased, using the same quantity of GPUs. We see a significant improvement in the tail latency variance.

Similarly, p90 time per token went from ~100 ms with high variance to 20-30 ms, a 3x improvement in intertoken latency.

Since agentic use cases usually have long contexts, we optimize for efficient prompt caching in order to not recompute input tensors on every turn. We leverage a header called x-session-affinity in order to help requests route to the right region that previously had the computed input tensors. We wrote about this in our original blog post about launching large LLMs on Workers AI. We added session affinity headers to popular agent harnesses like OpenCode, where we noticed a significant increase in total throughput. A small difference in prompt caching from our users can sum to a factor of additional GPUs needed to run a model. While we have KV-aware routing internally, we also rely on clients sending the x-session-affinity in order to be explicit about prompt caching. We incentivize the use of the header by offering discounted cached tokens. We highly encourage users to leverage prompt caching in order to have faster inference and cheaper pricing.

We worked with our heaviest internal users to adopt this header. The result was an increase in input token cache hit ratios from 60% to 80% during peak times. This significantly increases the request throughput that we can handle, while offering better performance for interactive or time-sensitive sessions like OpenCode or AI code reviews.

As we’re serving larger models now, one instance can span multiple GPUs. This means that we had to find an efficient way to share KV cache across GPUs. KV cache is where all the input tensors from prefill (result of prompts in a session) are stored, and initially lives in the VRAM of a GPU. Every GPU has a fixed VRAM size, but if your model instance requires multiple GPUs, there needs to be a way for the KV cache to live across GPUs and talk to each other. To achieve this for Kimi, we leveraged Moonshot AI’s Mooncake Transfer Engine and Mooncake Store.

Mooncake’s Transfer Engine is a high-performance data transfer framework. It works with different Remote Direct Memory Access (RDMA) protocols such as NVLink and NVMe over Fabric, which enables direct memory-to-memory data transfer without involving the CPU. It improves the speed of transferring data across multiple GPU machines, which is particularly important in multi-GPU and multi-node configurations for models. 

When paired with LMCache or SGLang HiCache, the cache is shared across all nodes in the cluster, allowing a prefill node to identify and re-use a cache from a previous request that was originally pre-filled on a different node. This eliminates the need for session aware routing within a cluster and allows us to load balance the traffic much more evenly. Mooncake Store also allows us to extend the cache beyond GPU VRAM, and leverage NVMe storage. This extends the time that sessions remain in cache, improving our cache hit ratio and allowing us to handle more traffic and offer better performance to users.

LLMs work by predicting the next token in a sequence, based on the tokens that came before it. With a naive implementation, models only predict the next n token, but we can actually make it predict the next n+1, n+2... tokens in a single forward pass of the model. This popular technique is known as speculative decoding, which we’ve written about in a previous post on Workers AI. 

With speculative decoding, we leverage a smaller LLM (the draft model) to generate a few candidate tokens for the target model to choose from. The target model then just has to select from a small pool of candidate tokens in a single forward pass. Validating the tokens is faster and less computationally expensive than using the larger target model to generate the tokens. However, quality is still upheld as the target model ultimately has to accept or reject the draft tokens.

In agentic use cases, speculative decoding really shines because of the volume of tool calls and structured outputs that models need to generate. A tool call is largely predictable — you know there will be a name, description, and it’s wrapped in a JSON envelope.

To do this with Kimi K2.5, we leverage NVIDIA’s EAGLE-3 (Extrapolation Algorithm for Greater Language-model Efficiency) draft model. The levers for tuning speculative decoding include the number of future tokens to generate. As a result, we’re able to achieve high-quality inference while speeding up tokens per second throughput.

As we announced during Birthday Week in 2025, Cloudflare has a proprietary inference engine, Infire, that makes machine learning models faster. Infire is an inference engine written in Rust, designed to support Cloudflare’s unique challenges with inference given our distributed global network. We’ve extended Infire support for this new class of large language models we are planning to run, which meant we had to build a few new features to make it all work.

Large language models like Kimi K2.5 are over 1 trillion parameters, which is about 560GB of model weights. A typical H100 has about 80GB of VRAM and the model weights need to be loaded in GPU memory in order to run. This means that a model like Kimi K2.5 needs at least 8 H100s in order to load the model into memory and run — and that’s not even including the extra VRAM you would need for KV Cache, which includes your context window.

Since we initially launched Infire, we had to add support for multi-GPU, letting the inference engine run across multiple GPUs in either pipeline-parallel or tensor-parallel modes with expert-parallelism supported as well.

For pipeline parallelism, Infire attempts to properly load balance all stages of the pipeline, in order to prevent the GPUs of one stage from starving while other stages are executing. On the other hand, for tensor parallelism, Infire optimizes for reducing cross-GPU communication, making it as fast as possible. For most models, utilizing both pipeline parallelism and tensor parallelism in tandem provides the best balance of throughput and latency.

While already having much lower GPU memory overhead than vLLM, we optimized Infire even further, tightening the memory required for internal state like activations. Currently Infire is capable of running Llama 4 Scout on just two H200 GPUs with more than 56 GiB remaining for KV-cache, sufficient for more than 1.2m tokens. Infire is also capable of running Kimi K2.5 on 8 H100 GPUs (yes that is H100), with more than 30 GiB still available for KV-cache. In both cases you would have trouble even booting vLLM in the first place.

While adding multi-GPU support, we identified additional opportunities to improve boot times. Even for the largest models, such as Kimi K2.5, Infire can begin serving requests in under 20 seconds. The load times are only bounded by the drive speed.

Investing in our proprietary inference engine enables us to maximize our hardware by getting up to 20% higher tokens per second throughput on unconstrained systems, and also enabling us to use lower-end hardware to run the latest models, where it was previously completely infeasible.

New technologies, research, and models come out on a weekly basis for the machine learning community. We’re continuously optimizing our technology stack in order to provide high-quality, performant inference for our customers while operating our GPUs efficiently. If these sound like interesting challenges for you – we’re hiring!

At Cloudflare, we have been building these primitives for years: Durable Objects for state persistence, Workflows for long running tasks, and Dynamic Workers or Sandbox containers for secure execution. Powerful abstractions like the Agents SDK are designed to help you build agents on top of Cloudflare’s Developer Platform.

But these primitives only provided the execution environment. The agent still needed a model capable of powering it. 

Starting today, Workers AI is officially in the big models game. We now offer frontier open-source models on our AI inference platform. We’re starting by releasing Moonshot AI’s Kimi K2.5 model on Workers AI. With a full 256k context window and support for multi-turn tool calling, vision inputs, and structured outputs, the Kimi K2.5 model is excellent for all kinds of agentic tasks. By bringing a frontier-scale model directly into the Cloudflare Developer Platform, we’re making it possible to run the entire agent lifecycle on a single, unified platform.

The heart of an agent is the AI model that powers it, and that model needs to be smart, with high reasoning capabilities and a large context window. Workers AI now runs those models.

We spent the last few weeks testing Kimi K2.5 as the engine for our internal development tools. Within our OpenCode environment, Cloudflare engineers use Kimi as a daily driver for agentic coding tasks. We have also integrated the model into our automated code review pipeline; you can see this in action via our public code review agent, Bonk, on Cloudflare GitHub repos. In production, the model has proven to be a fast, efficient alternative to larger proprietary models without sacrificing quality.

Serving Kimi K2.5 began as an experiment, but it quickly became critical after reviewing how the model performs and how cost-efficient it is. As an illustrative example: we have an agent that does security reviews of Cloudflare’s codebases. This agent processes over 7B tokens per day, and using Kimi, it has caught more than 15 confirmed issues in a single codebase. Doing some rough math, if we had run this agent on a mid-tier proprietary model, we would have spent $2.4M a year for this single use case, on a single codebase. Running this agent with Kimi K2.5 cost just a fraction of that: we cut costs by 77% simply by making the switch to Workers AI.

As AI adoption increases, we are seeing a fundamental shift not only in how engineering teams are operating, but how individuals are operating. It is becoming increasingly common for people to have a personal agent like OpenClaw running 24/7. The volume of inference is skyrocketing.

This new rise in personal and coding agents means that cost is no longer a secondary concern; it is the primary blocker to scaling. When every employee has multiple agents processing hundreds of thousands of tokens per hour, the math for proprietary models stops working. Enterprises will look to transition to open-source models that offer frontier-level reasoning without the proprietary price tag. Workers AI is here to facilitate this shift, providing everything from serverless endpoints for a personal agent to dedicated instances powering autonomous agents across an entire organization.

Workers AI has served models, including LLMs, since its launch two years ago, but we’ve historically prioritized smaller models. Part of the reason was that for some time, open-source LLMs fell far behind the models from frontier model labs. This changed with models like Kimi K2.5, but to serve this type of very large LLM, we had to make changes to our inference stack. We wanted to share with you some of what goes on behind the scenes to support a model like Kimi.

We’ve been working on custom kernels for Kimi K2.5 to optimize how we serve the model, which is built on top of our proprietary Infire inference engine. Custom kernels improve the model’s performance and GPU utilization, unlocking gains that would otherwise go unclaimed if you were just running the model out of the box. There are also multiple techniques and hardware configurations that can be leveraged to serve a large model. Developers typically use a combination of data, tensor, and expert parallelization techniques to optimize model performance. Strategies like disaggregated prefill are also important, in which you separate the prefill and generation stages onto different machines in order to get better throughput or higher GPU utilization. Implementing these techniques and incorporating them into the inference stack takes a lot of dedicated experience to get right. 

Workers AI has already done the experimentation with serving techniques to yield excellent throughput on Kimi K2.5. A lot of this does not come out of the box when you self-host an open-source model. The benefit of using a platform like Workers AI is that you don’t need to be a Machine Learning Engineer, a DevOps expert, or a Site Reliability Engineer to do the optimizations required to host it: we’ve already done the hard part, you just need to call an API.

In concert with this launch, we’ve also improved our platform and are releasing several new features to help you build better agents.

When you work with agents, you are likely sending a large number of input tokens as part of the context: this could be detailed system prompts, tool definitions, MCP server tools, or entire codebases. Inputs can be as large as the model context window, so in theory, you could be sending requests with almost 256k input tokens. That’s a lot of tokens.

When an LLM processes a request, the request is broken down into two stages: the prefill stage processes input tokens and the output stage generates output tokens. These stages are usually sequential, where input tokens have to be fully processed before you can generate output tokens. This means that sometimes the GPU is not fully utilized while the model is doing prefill.

With multi-turn conversations, when you send a new prompt, the client sends all the previous prompts, tools, and context from the session to the model as well. The delta between consecutive requests is usually just a few new lines of input; all the other context has already gone through the prefill stage during a previous request. This is where prefix caching helps. Instead of doing prefill on the entire request, we can cache the input tensors from a previous request, and only do prefill on the new input tokens. This saves a lot of time and compute from the prefill stage, which means a faster Time to First Token (TTFT) and a higher Tokens Per Second (TPS) throughput as you’re not blocked on prefill.

Workers AI has always done prefix caching, but we are now surfacing cached tokens as a usage metric and offering a discount on cached tokens compared to input tokens. (Pricing can be found on the model page.) We also have new techniques for you to leverage in order to get a higher prefix cache hit rate, reducing your costs.

In order to route to the same model instance and take advantage of prefix caching, we use a new x-session-affinity header. When you send this header, you’ll improve your cache hit ratio, leading to more cached tokens and subsequently, faster TTFT, TPS, and lower inference costs.

You can pass the new header like below, with a unique string per session or per agent. Some clients like OpenCode implement this automatically out of the box. Our Agents SDK starter has already set up the wiring to do this for you, too.

curl -X POST \
"https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/run/@cf/moonshotai/kimi-k2.5" \
  -H "Authorization: Bearer {API_TOKEN}" \
  -H "Content-Type: application/json" \
  -H "x-session-affinity: ses_12345678" \
  -d '{
    "messages": [
      {
        "role": "system",
        "content": "You are a helpful assistant."
      },
      {
        "role": "user",
        "content": "What is prefix caching and why does it matter?"
      }
    ],
    "max_tokens": 2400,
    "stream": true
  }'

Serverless inference is really hard. With a pay-per-token business model, it’s cheaper on a single request basis because you don’t need to pay for entire GPUs to service your requests. But there’s a trade-off: you have to contend with other people’s traffic and capacity constraints, and there’s no strict guarantee that your request will be processed. This is not unique to Workers AI — it’s evidently the case across serverless model providers, given the frequent news reports of overloaded providers and service disruptions. While we always strive to serve your request and have built-in autoscaling and rebalancing, there are hard limitations (like hardware) that make this a challenge.

For volumes of requests that would exceed synchronous rate limits, you can submit batches of inferences to be completed asynchronously. We’re introducing a revamped Asynchronous API, which means that for asynchronous use cases, you won’t run into Out of Capacity errors and inference will execute durably at some point. Our async API looks more like flex processing than a batch API, where we process requests in the async queue as long as we have headroom in our model instances. With internal testing, our async requests usually execute within 5 minutes, but this will depend on what live traffic looks like. As we bring Kimi to the public, we will tune our scaling accordingly, but the async API is the best way to make sure you don’t run into capacity errors in durable workflows. This is perfect for use cases that are not real-time, such as code scanning agents or research agents.

Workers AI previously had an asynchronous API, but we’ve recently revamped the systems under the hood. We now rely on a pull-based system versus the historical push-based system, allowing us to pull in queued requests as soon as we have capacity. We’ve also added better controls to tune the throughput of async requests, monitoring GPU utilization in real-time and pulling in async requests when utilization is low, so that critical synchronous requests get priority while still processing asynchronous requests efficiently.

To use the asynchronous API, you would send your requests as seen below. We also have a way to set up event notifications so that you can know when the inference is complete instead of polling for the request. 

// (1.) Push a request in queue
// pass queueRequest: true
let res = await env.AI.run("@cf/moonshotai/kimi-k2.5", {
  "requests": [{
    "messages": [{
      "role": "user",
      "content": "Tell me a joke"
    }]
  }, {
    "messages": [{
      "role": "user",
      "content": "Explain the Pythagoras theorem"
    }]
  }, ...{<add more requests in a batch>} ];
}, {
  queueRequest: true,
});


// (2.) grab the request id
let request_id;
if(res && res.request_id){
  request_id = res.request_id;
}
// (3.) poll the status
let res = await env.AI.run("@cf/moonshotai/kimi-k2.5", {
  request_id: request_id
});

if(res && res.status === "queued" || res.status === "running") {
 // retry by polling again
 ...
}
else 
 return Response.json(res); // This will contain the final completed response 

Get started with Kimi K2.5 on Workers AI today. You can read our developer docs to find out model information and pricing, and how to take advantage of prompt caching via session affinity headers and asynchronous API. The Agents SDK starter also now uses Kimi K2.5 as its default model. You can also connect to Kimi K2.5 on Workers AI via Opencode. For a live demo, try it in our playground.

And if this set of problems around serverless inference, ML optimizations, and GPU infrastructure sound  interesting to you — we’re hiring!