Opened a PR for Llama4 images to work for the AI SDK https://github.com/cloudflare/ai/pull/303 This

Opened a PR for Llama4 images to work for the AI SDK https://github.com/cloudflare/ai/pull/303
This follows the documentation found herehttps://developers.cloudflare.com/workers-ai/models/llama-4-scout-17b-16e-instruct/

GitHub

Add support for base64 images in chat messages by celestial-rose ·...

Handle base64 image data for Llama 4 in chat messages.
As per the documentation found here https://developers.cloudflare.com/workers-ai/models/llama-4-scout-17b-16e-instruct/

Cloudflare Docs

llama-4-scout-17b-16e-instruct

Meta's Llama 4 Scout is a 17 billion parameter model with 16 experts that is natively multimodal. These models leverage a mixture-of-experts architecture to offer industry-leading performance in text and image understanding.

CCelestial Rose Opened a PR for Llama4 images to work for the AI SDK https://github.com/cloudfla...

Celestial RoseOP•11/8/25, 9:55 PM

forgot i had to also change the "text" case so it outputs an object for it to work. put it to draft; wondering how to adapt without breaking anything. the inputs needs to have "type": "text""type": "text"

Andrew Hansen•11/9/25, 3:53 AM

Is there a plan to make GPU’s available to host custom models on someday? We have a custom model we are integrating currently

Tim•11/9/25, 9:49 PM

Morning from the land down under.

I'm attempting to use Flux again (I was using Deepgrams) - I am noticing their platform is more performative.

But, I'm having an issue where Flux appears to die. I'm not receiving any errors - but the Cloudflare instance accepts a couple of audio chunks, then appears to close the connection.

Here's the request ID: Flux connection confirmed { requestId: '616f1c9e-ab87-4942-9992-1a23dac29cab' }

Here's some logs, to see what happened:

[DEBUG][SpeechToTextOrchestrator] Handling end_of_turn { turnIndex: 0, transcript: 'Hi. How are you?' }
[DEBUG][TranscriptBuffer] No buffered content, content unchanged
[DEBUG][SpeechToTextOrchestrator] Processing new request { reason: 'first turn' }
[DEBUG][PendingMessageReconciler] Starting reconciliation { pendingMessagesCount: 1, newTranscript: 'Hi. How are you?' }
[DEBUG][SpeechToTextOrchestrator] Reconciled turn info {
originalTranscript: 'Hi. How are you?',
originalWordsCount: 4,
reconciledTranscript: 'Hi. How are you?',
reconciledWordsCount: 4,
pendingMessagesCount: 1
}
[DEBUG][TranscriptBuffer] Buffered turn info { turnIndex: 0, transcript: 'Hi. How are you?' }

[DEBUG][LangModelOrchestrator] Language model request completed for 1762724680932_bwfrjit { contentLength: 131, toolCallCount: 0 }
[DEBUG][TranscriptBuffer] Clearing buffer
[DEBUG][SpeechToTextOrchestrator] Handling end_of_turn { turnIndex: 1, transcript: 'Can you count to ten?' }
[DEBUG][TranscriptBuffer] No buffered content, content unchanged
[DEBUG][SpeechToTextOrchestrator] Processing new request { reason: 'first turn' }
[DEBUG][TranscriptBuffer] Clearing buffer
[INFO][RequestManager] Created new request 1762724692649_tfpk167
[DEBUG][CloudflareFlux] Flux WebSocket connection closed

This conenction closed at the end, was not me hanging up. It just did that itself. I don't have any errors or exceptions showing.

CCelestial Rose forgot i had to also change the "text" case so it outputs an object for it to wo...

Celestial RoseOP•11/10/25, 2:36 PM

perhaps even the type text should be default if no type is set

Celestial RoseOP•11/10/25, 3:51 PM

is it planned to add a AI Search (AutoRAG) parameter to the createWorkersAI of the Vercel AI SDK
https://developers.cloudflare.com/workers-ai/configuration/ai-sdk/
so we can init the binding with a autorag param ?

Cloudflare Docs

Vercel AI SDK

Workers AI can be used with the Vercel AI SDK for JavaScript and TypeScript codebases.

Celestial RoseOP•11/10/25, 3:54 PM

the same way we have "gateway"

Celestial RoseOP•11/10/25, 3:58 PM

oh wow, just found in the package a "createAutoRAG" instead of createWorkersAI which is a completely undocumented api, !

Celestial RoseOP•11/10/25, 4:09 PM

hmm its a bit weird, the generation model is hardcoded and is not using the model in the options, is it a non-released API on purpose ?

CCelestial Rose forgot i had to also change the "text" case so it outputs an object for it to wo...

Celestial RoseOP•11/10/25, 4:43 PM

funnily enough, returning an object instead of a string made me discover that it breaks the const query = messages.map(({ content, role }) => const query = messages.map(({ content, role }) => ${role}: ${content}).join("\n\n");).join("\n\n"); helper for the AutoRAG prompt system. also this is repeated itself twice and deserve a function on its own. will fix the PR for the base64 images and add it all there

Fernando•11/11/25, 8:03 AM

Deepseek OCR please

Celestial RoseOP•11/11/25, 8:21 AM

https://tenor.com/view/what-huh-confused-john-travolta-gif-16946746

Tenor

Celestial RoseOP•11/12/25, 8:58 AM

Hey ! Would be nice to add to the "Capability" to the Models page to filter by multimodal, image input, OpenAI Repsonses API etc etc

Capture_decran_2025-11-12_a_10.58.14.png

Celestial RoseOP•11/12/25, 1:17 PM

There's a bug in the autorag AI Binding where the totalUsage is always zero in stream mode when we reach the "finish" part

Celestial RoseOP•11/12/25, 1:17 PM

it does work with regular AI binding

blah•11/12/25, 3:02 PM

Requesting help with my community help request: "Request AI Streaming Binding Access". Can't post the link :(.

Issue is that this toy request (as well as my real ones) fail to return stream body:

export default {
async fetch(req, env) {
const model = “@cf/meta/llama-2-7b-chat-int8”;
const response = await env.DOCTALK_AI.run(model, {
prompt: “What is a tomato?”,
stream: true,
});
console.log(“response:”, response);
return new Response(response.body ? “

Stream OK” : “

No stream body”, {
status: 200,
});
},
};

No stream body
“[Worker] AI error: Error: No stream body returned”

any assistance greatly appreciated!!!

Bblah Requesting help with my community help request: "Request AI Streaming Binding Ac...

Celestial RoseOP•11/12/25, 5:50 PM

You shouldn’t do response.body
Streams don’t use body

Bblah Requesting help with my community help request: "Request AI Streaming Binding Ac...

Celestial RoseOP•11/12/25, 5:50 PM

heres the doc https://developers.cloudflare.com/workers-ai/configuration/bindings/#methods

Cloudflare Docs

Workers Bindings

Workers provides a serverless execution environment that allows you to create new applications or augment existing ones.

CCelestial Rose heres the doc https://developers.cloudflare.com/workers-ai/configuration/binding...

blah•11/12/25, 6:20 PM

Thank you!

Paul•11/13/25, 4:21 PM

Hi, wanted to ask the community here if there is a combination you find works well together of a Agentic Code CLI and a Model from Workers AI specifically. If your familiar with Claude Code, that's what I'm trying to determine if I could build a lower cost cloudflare based version of. Ideas?

PPaul Hi, wanted to ask the community here if there is a combination you find works we...

Paul•11/13/25, 4:25 PM

For example I tried combining Aider CLI with Amazon Nova (Not on cloudflare, just an example), and now I'm wonder if I can do the same with Cloudflare hotsed models? - BTW the Aider/Nova combination isn't working too well for me so far

PhoenixSpace•11/13/25, 4:25 PM

IME the cloudflare hosted models tend to be a bit more expensive then other providers. If you want low latency/edge usage, it's great, but if pure cost is your benchmark then other providers will get you a bit better pricing

PPhoenixSpace IME the cloudflare hosted models tend to be a bit more expensive then other prov...

Paul•11/13/25, 4:25 PM

Ok, thanks for the insight. I guess I didn't do a cost analysis, for some reason I sort of assumed they would be cheaper on cloudflare

like most stuff

PhoenixSpace•11/13/25, 4:28 PM

If you need to scale and run AI stuff globally/on the edge, their pricing is fantastic. If it’s just a single user internal tool, there are slightly cheaper options

Paul•11/13/25, 4:29 PM

I need to get my head around "Neurons" pricing - so I can comapre apples to apples with other solutions.

PhoenixSpace•11/13/25, 4:30 PM

The model pages usually have price per M tokens

PhoenixSpace•11/13/25, 4:30 PM

https://developers.cloudflare.com/workers-ai/models/gpt-oss-120b/

Cloudflare Docs

gpt-oss-120b

OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases – gpt-oss-120b is for production, general purpose, high reasoning use-cases.

Brick_Syntax•11/13/25, 9:52 PM

Hey! so According to chat gpt I can use worker AIs to make my self a AI api, but it says there is limit, it said something about 10,00 neurons per day, now I wanted a better source so is this true? and if so how many neurons per task?

boke0•11/15/25, 12:45 AM

Hi there, I could use some help.
A few days ago, my product, which is built on Cloudflare Workers and integrates with Workers AI (GPT-OSS-120b) through bindings, broke because the interface changed (prompt_tokensprompt_tokens and completion_tokenscompletion_tokens were replaced with input_tokensinput_tokens and output_tokensoutput_tokens).
And today, I ran into another issue: the instructionsinstructions parameter stopped working when the input was in an OpenAI-like format (like [{role: "assistant", content: "foobar" }, ...][{role: "assistant", content: "foobar" }, ...]).

Where can I check for these kinds of interface changes in advance so I can update my product before it breaks?

bayes•11/15/25, 8:41 PM

did someone tested deepgram's nova-3 in workers ai with diarization? always getting timeout regardless of file size or type

KZ•11/16/25, 8:13 PM

is there a channel that has the ball on web-bot-auth specifically the http message signature (9421) is that this group or agents or some other?

samjs•11/17/25, 3:22 PM

https://blog.cloudflare.com/replicate-joins-cloudflare/

The Cloudflare Blog

Replicate is joining Cloudflare

Bringing Replicate’s tools into Cloudflare will continue to make our Workers Platform the best place on the Internet to build and deploy any AI or agentic workflow.

Celestial RoseOP•11/17/25, 3:29 PM

Holy s*** the Replicate news, CONGRATS !!! Yay to us

Ssamjs https://blog.cloudflare.com/replicate-joins-cloudflare/

LordSilver•11/18/25, 2:32 AM

So... Workers AI = Replicate now? Will you be able to set the GPU model?

LordSilver•11/18/25, 2:33 AM

I've never checked Workers AI vs Replicate pricing for well-known models.

PPhoenixSpace IME the cloudflare hosted models tend to be a bit more expensive then other prov...

zsueo•11/18/25, 11:49 AM

What cheaper options are you referring to, cloudflares paid worker plan is the cheapest I've seen so far

Zzsueo What cheaper options are you referring to, cloudflares paid worker plan is the c...

PhoenixSpace•11/18/25, 1:05 PM

For LLM inference?

PPhoenixSpace For LLM inference?

zsueo•11/18/25, 1:20 PM

Yeah

Zzsueo Yeah

PhoenixSpace•11/18/25, 2:18 PM

Google vertex and Amazon bedrock running the same model is usually a bit cheaper, same with fireworks AI and deepinfra

PhoenixSpace•11/18/25, 2:19 PM

Similar in pricing is groq (not to be confused with grok!) but faster output

PhoenixSpace•11/18/25, 2:19 PM

This is why I’m stoked about Replicate, I think this is gonna be huge

Original message was deleted

QuentinCody•11/18/25, 2:30 PM

Yes, I'm working on a chatbot+MCP Client that comes pre-loaded with custom MCP Servers I've built

Vivek_m4_•11/20/25, 6:06 AM

Hi, I built a huge AI tool using Cloudflare Workers (Spider AI).
I want to discuss potential collaboration or GPU credits.
Whom should I contact from the Cloudflare team?

Bblah Requesting help with my community help request: "Request AI Streaming Binding Ac...

Vivek_m4_•11/20/25, 6:19 AM

export default {
async fetch(req, env) {
const model = "@cf/meta/llama-2-7b-chat-int8";

const stream = await env.DOCTALK_AI.run(model, {
prompt: "What is a tomato?",
stream: true,
});

return new Response(stream, {
headers: {
"Content-Type": "text/event-stream",
},
});
},
};
Hello bro this is correct code

vcc•11/21/25, 7:37 AM

fetch(`https://api.cloudflare.com/client/v4/accounts/${this.env.CF_ACCOUNT}/ai/run/@cf/deepgram/nova-3?${new URLSearchParams({
            encoding: 'linear16',
            sample_rate: '16000',
                        language:  ?
}).toString()}`)

fetch(`https://api.cloudflare.com/client/v4/accounts/${this.env.CF_ACCOUNT}/ai/run/@cf/deepgram/nova-3?${new URLSearchParams({
            encoding: 'linear16',
            sample_rate: '16000',
                        language:  ?
}).toString()}`)

Can I specify a language for it? and how many languages support it?

VVivek_m4_export default { async fetch(req, env) { const model = "@cf/meta/llama-2-7...

blah•11/21/25, 5:43 PM

Thanks Vivek

Ahmed bat•11/21/25, 6:09 PM

Image models return 0 bytes but text models work

Ahmed bat•11/21/25, 6:10 PM

hey there im trying to use leonardo ai to create images from cloudflare workers ai ... i ve tested it on my n8n workflow it wroks good with other models but with leonardo models i ve got thiserror : Image models return 0 bytes but text models work

Catkin038•11/22/25, 10:24 AM

Hello everyone, my Workers AI returned Error 1101, and the reason in the statistics is Uncaught Exception. How should I troubleshoot this?

Catkin038•11/22/25, 10:39 AM

I found the problem, 5028: This model was deprecated on 2025-10-01. Please use an alternative model.

Catkin038•11/22/25, 10:42 AM

I need to confirm whether @cf/qwen/qwen2.5-coder-32b-instruct still accepts support.

Opened a PR for Llama4 images to work for the AI SDK https://github.com/cloudflare/ai/pull/303 This

Similar Threads