Hey @Fabian . Streaming for the gpt-oss models is not currently supported, but we're working on it a

S

samjsOP•8/27/25, 4:48 PM

Hey folks! Happy AI Week. I'm on the Workers AI team and will be around to help answer any questions you all have

HHimanshu Is there an ETA for function calling support for gpt oss models in workers ai?

S

samjsOP•8/27/25, 4:52 PM

Hey @Himanshu ! Function calling should work with the gpt-oss models. Here's an example of that with our sandbox SDK: https://github.com/cloudflare/sandbox-sdk/blob/main/examples/code-interpreter/src/index.ts#L73-L91

We've also tested against the compatibilty suite here: https://github.com/openai/gpt-oss/tree/main/compatibility-test and are getting ~93% successes. Do you have an example of what you were trying that didn't work?

We're also actively working on integrating support for the builtin

code_interpreter

code_interpreter

tool to make some of this even easier.

MMirza Hey James, I believe you're from the cloudflare team. I wanted to know why are...

S

samjsOP•8/27/25, 4:56 PM

Hi @Mirza

currently we've mostly been running open source models on our platform. Though as of today we have new partner models we're supporting as James linked above.

Definitely feel free to share your feedback in your office hours today. Sharing here is also a good way for us to know what is important to folks. Are you open to sharing more about your use case? I'd also love to hear if you've tried the gpt-oss models for your use case and how those compare.

Ssamjs Hi @Mirza 👋 currently we've mostly been running open source models on our pla...

M

Mirza•8/27/25, 4:58 PM

Yeah, let me DM you as I'm pre-occupied today so won't be able to attend the office hours ig.

H

Himanshu•8/27/25, 5:01 PM

Getting this error when trying out deepgram/nova-3:

{
"message": "Error: Network connection lost.",
"code": 6001
}

HHimanshu Getting this error when trying out deepgram/nova-3: { "mess...

S

samjsOP•8/27/25, 5:02 PM

Taking a look on our side

Ssamjs Hey @Himanshu ! Function calling should work with the gpt-oss models. Here's an ...

H

Himanshu•8/27/25, 5:05 PM

Amazing! I was trying it out with ai sdk, but the normal tool call invocation that worked with other function calling models like Llama 3.3, didn't work. I checked the docs, and there was no mention of tool call support for these models, so I took a back seat. I'll retry and update you. Thanks again.

H

Himanshu•8/27/25, 5:44 PM

Leonardo lucid-origin is a fantastic model!!

Prompt: Ultra-realistic 3D render of Times Square at night with neon lights and bustling atmosphere. A massive digital billboard dominates the scene, displaying the Cloudflare logo with the words ‘Workers AI’ in bold white font on a vibrant orange background. The billboard is crisp, glowing brightly against the dark sky, reflecting on nearby glass buildings. Camera shot captured from across the street at a slightly low angle, showing street lights, cars, and people walking, adding scale and realism. Photorealistic textures, cinematic lighting, and a dramatic depth-of-field emphasize the billboard as the centerpiece.

S

samjsOP•8/27/25, 6:08 PM

Love it! Thanks for sharing

HHimanshu Getting this error when trying out deepgram/nova-3: { "mess...

S

samjsOP•8/27/25, 8:33 PM

FYI we have a fix going out for this issue shortly

A

akazwz•8/28/25, 11:04 AM

Workers AI released so many models, which is exciting! Some models support WebSocket, which is awesome, but I can't find any documentation on how to use them with WebSocket. Can anyone help?

J

John Spurlock•8/28/25, 2:09 PM

pricing for the deepgram nova models should be based on audio input, right? Here is a quick PR to tweak the pricing page wording https://github.com/cloudflare/cloudflare-docs/pull/24763

GitHub

Prices should be per input for audio input models by johnspurlock-s...

Aakazwz Workers AI released so many models, which is exciting! Some models support WebSo...

S

samjsOP•8/28/25, 2:55 PM

I think we're still working on the documentation! But in short it should look something like:

export default {
  async fetch(request, env, ctx): Promise<Response> {
    const resp = await env.AI.run("@cf/deepgram/aura-1", {}, {
      websocket: true
    });
    return resp;
  },
} satisfies ExportedHandler<Env>;

export default {
  async fetch(request, env, ctx): Promise<Response> {
    const resp = await env.AI.run("@cf/deepgram/aura-1", {}, {
      websocket: true
    });
    return resp;
  },
} satisfies ExportedHandler<Env>;

export default {
  async fetch(request, env, ctx): Promise<Response> {
    const resp = await env.AI.run("@cf/deepgram/aura-1", {}, {
      websocket: true
    });
    return resp;
  },
} satisfies ExportedHandler<Env>;

export default {
  async fetch(request, env, ctx): Promise<Response> {
    const resp = await env.AI.run("@cf/deepgram/aura-1", {}, {
      websocket: true
    });
    return resp;
  },
} satisfies ExportedHandler<Env>;

JJohn Spurlock pricing for the deepgram nova models should be based on audio *input*, right? H...

S

samjsOP•8/28/25, 3:01 PM

Thanks for flagging! Checking with the team, but yes that also sounds right to me given that it's an audio -> text model

Ssamjs I think we're still working on the documentation! But in short it should look so...

J

John Spurlock•8/29/25, 4:17 PM

thanks this is helpful for bindings. What is the equivalent way of using websockets over the CF REST api? https://developers.cloudflare.com/api/resources/ai/methods/run/

J

John Spurlock•8/29/25, 4:27 PM

how does the cf partnership thing work with deepgram in terms of the "model improvement plan" - does setting

mip_opt_out=true

mip_opt_out=true

affect the unit pricing stated on the cf api page? (when using deepgram directly that significantly increases the price)

https://developers.cloudflare.com/workers-ai/models/nova-3/

Ssamjs FYI we have a fix going out for this issue shortly

H

Himanshu•8/30/25, 5:48 PM

Is nova-3 operational now? I'm still getting the same error as before in the API.

A

Ana•8/31/25, 4:45 PM

Hello. I’m experimenting with Traditional function calling and it seems to only work with one of the 4 models that should support it:

@hf/nousresearch/hermes-2-pro-mistral-7b

@hf/nousresearch/hermes-2-pro-mistral-7b

For the other models I get the response:

 {
  "errors": [
    {
      "message": "AiError: AiError: Invalid input",
      "code": 8001
    }
  ],
  "success": false,
  "result": {},
  "messages": []
}

 {
  "errors": [
    {
      "message": "AiError: AiError: Invalid input",
      "code": 8001
    }
  ],
  "success": false,
  "result": {},
  "messages": []
}

 {
  "errors": [
    {
      "message": "AiError: AiError: Invalid input",
      "code": 8001
    }
  ],
  "success": false,
  "result": {},
  "messages": []
}

 {
  "errors": [
    {
      "message": "AiError: AiError: Invalid input",
      "code": 8001
    }
  ],
  "success": false,
  "result": {},
  "messages": []
}

To stop receiving this message I need to give a type: “function” to my tools and wrap them in a function:

tools: [
    {
      type: “function”,
      function: { 
        name: "getWeather",
        description: "Return the weather for a latitude and longitude",
        parameters: {
          // ....
        },
      },
    }],

tools: [
    {
      type: “function”,
      function: { 
        name: "getWeather",
        description: "Return the weather for a latitude and longitude",
        parameters: {
          // ....
        },
      },
    }],

tools: [
    {
      type: “function”,
      function: { 
        name: "getWeather",
        description: "Return the weather for a latitude and longitude",
        parameters: {
          // ....
        },
      },
    }],

tools: [
    {
      type: “function”,
      function: { 
        name: "getWeather",
        description: "Return the weather for a latitude and longitude",
        parameters: {
          // ....
        },
      },
    }],

But this way none of the other 3 models (

@cf/meta/llama-3.3-70b-instruct-fp8-fast

@cf/meta/llama-3.3-70b-instruct-fp8-fast

,

@cf/mistralai/mistral-small-3.1-24b-instruct

@cf/mistralai/mistral-small-3.1-24b-instruct

,

@cf/meta/llama-4-scout-17b-16e-instruct)

@cf/meta/llama-4-scout-17b-16e-instruct)

can use function calling at all (the reply is always text).

Anyone else with the same issue?

J

joe•8/31/25, 8:11 PM

are there any plans to expand the options that can be passed into runWithTools? not being able to set max output tokens is a real pain

AAna Hello. I’m experimenting with Traditional function calling and it seems to only ...

J

joe•8/31/25, 8:27 PM

I can also reproduce this

AAna Hello. I’m experimenting with Traditional function calling and it seems to only ...

P

Proman4713 🇵🇸•9/1/25, 9:50 AM

i have the same issue

B

bpence•9/1/25, 5:11 PM

Is there anyone here that could help with a whisper ai issue? I've got the full flow working (r2/whisper/kv) but the transcript text is really bad with lots of repetition of sentences and cut-off words. I use whisper on my local GPU and it's nearly perfect, but whisper on CF is pretty bad no matter which model I use.

PProman4713 🇵🇸i have the same issue

S

samjsOP•9/3/25, 12:43 PM

Sorry for the delay folks. I'm out of office this week but will look into this next week if none of my colleagues get to it first

Ssamjs Sorry for the delay folks. I'm out of office this week but will look into this n...

P

Proman4713 🇵🇸•9/3/25, 2:20 PM

Alright, I can work on other things until then

AAna Hello. I’m experimenting with Traditional function calling and it seems to only ...

Z

zsueo•9/3/25, 11:58 PM

huh i can get function calling to work but for llama instruct it sometimes just hangs without giving a response

Z

zsueo•9/3/25, 11:58 PM

id say it works around 30% of the time

Z

zsueo•9/3/25, 11:59 PM

it also likes to leak the function call into plain text

Z

zsueo•9/4/25, 12:11 AM

is the api docs for the traditional tool calling out of date?

J

jd•9/4/25, 5:54 AM

im super excited about workers ai https://blog.cloudflare.com/workers-ai/

and curious what's the latest status on it? im primarily looking into deploying my custom python script + model with compilation graph cache and see if cloudflare is able to run it at global scale serverless. with great cold start time.

The Cloudflare Blog

Workers AI: serverless GPU-powered inference on Cloudflare’s glob...

We are excited to launch Workers AI - an AI inference as a service platform, empowering developers to run AI models with just a few lines of code, all powered by our global network of GPUs

J

jd•9/4/25, 5:55 AM

for models im mean tasks like

1) upscaling
2) voice gen
3) ad-hoc AI processing
4) or even large diffusion transformer models that could take minimum of H100 to run if that's on the roadmap

Zzsueo is the api docs for the traditional tool calling out of date?

Z

zsueo•9/4/25, 10:04 AM

actually is there documentation on how the ai is supposed to return tool calls?

Y

yureckey•9/4/25, 10:12 AM

I'm testing aura-1 text2speech. There is no mentioning of Text input limit, but it seems to have 2k symbols limit (as I understand Deepgram have this limit). When the text is bigger - there is no proper error message, please fix (and add to doc)

Zzsueo is the api docs for the traditional tool calling out of date?

A

Ana•9/4/25, 11:56 AM

Probably, though it works for the hermes model. https://developers.cloudflare.com/workers-ai/features/function-calling/

AAna Probably, though it works for the hermes model. https://developers.cloudflare.co...

Z

zsueo•9/4/25, 12:03 PM

thanks! though i think the problem im facing is that A) my code was buggy and B) the input schema changed

Z

zsueo•9/4/25, 12:03 PM

took me a while to figure that out tho

SSunil Pai folks, we just published v2.0.0 of workers-ai-provider, the main feature of whic...

P

Proman4713 🇵🇸•9/4/25, 5:34 PM

does this fix the issue we were talking about?

SSunil Pai sorry, not sure what issue you're referring to

P

Proman4713 🇵🇸•9/4/25, 5:34 PM

AiError: Invalid Input 8081 when trying to use traditional function calling

P

Proman4713 🇵🇸•9/4/25, 5:35 PM

the response from the AI request (decoded from base64 after setting log-level to debug and turning off sanitisation):

{"httpCode":400,"internalCode":8001,"message":"AiError: AiError: Invalid input (95e650b5-30fd-461e-a800-95d18c5295ef)","name":"AiError","skipSentry":true,"description":"Invalid input"}

{"httpCode":400,"internalCode":8001,"message":"AiError: AiError: Invalid input (95e650b5-30fd-461e-a800-95d18c5295ef)","name":"AiError","skipSentry":true,"description":"Invalid input"}

{"httpCode":400,"internalCode":8001,"message":"AiError: AiError: Invalid input (95e650b5-30fd-461e-a800-95d18c5295ef)","name":"AiError","skipSentry":true,"description":"Invalid input"}

{"httpCode":400,"internalCode":8001,"message":"AiError: AiError: Invalid input (95e650b5-30fd-461e-a800-95d18c5295ef)","name":"AiError","skipSentry":true,"description":"Invalid input"}

SSunil Pai just saw it. this wasn't related to that, this is the library on top of workers-...

P

Proman4713 🇵🇸•9/4/25, 5:36 PM

oh okay, but still this issue works whenever i add

tools

tools

to my AI.run function

P

Proman4713 🇵🇸•9/4/25, 5:37 PM

i was in the wrong channel, but these are the details: coding-help

SSunil Pai just saw it. this wasn't related to that, this is the library on top of workers-...

P

Proman4713 🇵🇸•9/4/25, 6:32 PM

uhhh sorry i meant "this issue happens" not works

P

Proman4713 🇵🇸•9/4/25, 6:32 PM

it doesnt work

P

Proman4713 🇵🇸•9/4/25, 6:32 PM

it shows this error whenever i try to use traditional function calling

S

starcarlos.•9/5/25, 3:00 AM

When I use worker ai, I get an error.
429 , reason: Too Many Requests,text:{"errors":[{"message":"AiError: AiError: Capacity temporarily exceeded, please try again.

M

maj•9/5/25, 5:52 AM

I'm trying to use AutoRAG to spell check names. I have a list of names in a R2 bucket (split in ~3MB chunks). AutoRAG claims it finished processing them. I then use

c.env.AI.autorag('my-rag-name').aiSearch()

c.env.AI.autorag('my-rag-name').aiSearch()

c.env.AI.autorag('my-rag-name').aiSearch()

c.env.AI.autorag('my-rag-name').aiSearch() to ask it basically "give me the correct version of the given name". Hoping it would fix the spelling of "Suma Covjek" to "Šuma Čovjek", as listed in one of the files. But results are underwhelming. When I looked at the AI Gateway and what the

aiSearch()

aiSearch()

aiSearch()

aiSearch() used for the default system message, I see "...no relevant documents were retrieved for the user's query...". For each and all of them. If I do a search in the playground it would sometimes return a file, but often not the right or relevant one, even if the exact correct spelling is given. What am I missing? How come

aiSearch()

aiSearch()

aiSearch()

aiSearch() would not find my content?

Hey @Fabian . Streaming for the gpt-oss models is not currently supported, but we're working on it a

Similar Threads

Similar Threads

Similar Threads