Naspers and Prosus launch free AI platform for South African businesses that offers Claude

[Luis]

Naspers and Prosus launch free AI platform for South African businesses that offers Claude

Naspers, through its subsidiary Prosus, launched a new agentic AI platform for businesses in South Africa on Tuesday, called ToqanClaw, that is completely free to use.

Similar to the Open Claw platform, ToqanClaw draws on multiple AI models from major providers, including Anthropic and other open-source models.

 

[The_Ogre]

There's a snowball's chance in hell Anthropic would zero rate token usage in a schithole country like ours.

 

[quovadis]

What’s the catch?

 

[Fulcrum29]

There's a snowball's chance in hell Anthropic would zero rate token usage in a schithole country like ours.

All you need to know is here:

Terms and Conditions | Toqan

The terms governing your access to and use of the Toqan Beta Trial.

toqan.ai

 

[rustypup]

What’s the catch?

You're giving away your IP?

Like all the other "free" SKUs everywhere always

 

[quovadis]

You're giving away your IP?

Like all the other "free" SKUs everywhere always

Fun.

License to Supplier. By submitting Customer Content, you grant us a worldwide, royalty-free, and sublicensable license to use, host, and reproduce your Customer Content solely to operate, maintain, and improve the Services (including fine tuning of our proprietary models) and to comply with legal obligations. We will not use your content to train any third-party large language models.

Feedback. If you provide suggestions or feedback (“Feedback”), you grant us an irrevocable, perpetual, and royalty-free license to use and implement that Feedback for any purpose without any obligation or compensation to you.

 

[FiestaST]

AI for the Real Economy: Prosus introduces Zapia and ToqanClaw | Toqan

Prosus will push the OpenClaw inspired tools to its 5 million partners and 1 billion customers.

toqan.ai

 

[NarrowBandFtw]

I give it 6 months. Bait & switch.
No such thing as a free lunch.

which is exactly why it's vital to invest one's time in tools that can switch at the drop of a hat

tool X (Hermes in my case), API key Y that has a nice promotion or free tier or whatever it may be

Y can change tomorrow and I would not give a flying 'F', I persist with X and pick a new Y

 

[NarrowBandFtw]

I'm giving Hermes another 6 months before I dip my toes in...

https://medium.com/@sathishkraju/i-switched-from-openclaw-to-hermes-agent-heres-what-nobody-told-me-5f33a746b6ca

article is full of half truths fyi

Self-learning is disabled by default

false, I've never enabled it and when it completes a complex task or I point out mistakes it updates its own memory and skill files

Not a code-generation tool

I'd LOVE to see by what metric this guy reckons other tools "outperform it" in coding tasks, it literally has a Claude skill and a whole host of coding skills, as long as you connect it to a capable model it is on par with everything else for coding

Memory is opaque

not sure who will ever care about this TBH, but Hermes has a MEMORY.md it updates too, wtf is this SQLite claim?

Fewer integrations — and that’s a deliberate choice

100% correct, but the examples listed are a joke:

For users who need WhatsApp + Telegram + Slack + Discord

Hermes has all of those ...

Young project, fewer releases

Hermes doesn't really have "releases", this is a nonsensical point about version numbering schemes chosen by different projects, at any given moment you can run 'hermes update' and you'll see 100+ github commits every few hours being updated, it is more active than most people are going to be able to keep up with

 

[Sapphiron]

What’s the catch?

Yes, Who's venture capital are they burning and why? Maybe they are using it to funnel revenue to a 3rd party entity to write it off as a loss on the main business.

 

[Fulcrum29]

Fun.

License to Supplier. By submitting Customer Content, you grant us a worldwide, royalty-free, and sublicensable license to use, host, and reproduce your Customer Content solely to operate, maintain, and improve the Services (including fine tuning of our proprietary models) and to comply with legal obligations. We will not use your content to train any third-party large language models.

Feedback. If you provide suggestions or feedback (“Feedback”), you grant us an irrevocable, perpetual, and royalty-free license to use and implement that Feedback for any purpose without any obligation or compensation to you.

Hey... at least you own it,

Your Ownership. You retain all ownership rights in your Customer Content.

Output. Intellectual property created by you using the product is owned by you, subject to Applicable Law.

Yeah... also,

Product Evolution. As this is a free trial, we reserve the right to amend these Terms and to modify, suspend, or terminate any Product features or the Services entirely at any time, without liability to you. Any amendment to these Terms is effective immediately upon publishing the updated Terms on the Product platform and/or website. Your continued use of the Services after the effective date of any amendment constitutes acceptance of the amended Terms.

Basically, it is their sandbox and this is an open invitation to chuck your toys into it. Just know that even when your toys leave the sandbox, that they have a right to those toys. Whilst this is still an open invitation, the host may or may not turn hostile.

All that said, it is pretty normal with all these services. The golden rule! For any enterprise, whether small or large, don't share your IP, ideas, and anything considered secret or proprietary.

 

[Fulcrum29]

article is full of half truths fyi


false, I've never enabled it and when it completes a complex task or I point out mistakes it updates its own memory and skill files


I'd LOVE to see by what metric this guy reckons other tools "outperform it" in coding tasks, it literally has a Claude skill and a whole host of coding skills, as long as you connect it to a capable model it is on par with everything else for coding


not sure who will ever care about this TBH, but Hermes has a MEMORY.md it updates too, wtf is this SQLite claim?


100% correct, but the examples listed are a joke:

Hermes has all of those ...


Hermes doesn't really have "releases", this is a nonsensical point about version numbering schemes chosen by different projects, at any given moment you can run 'hermes update' and you'll see 100+ github commits every few hours being updated, it is more active than most people are going to be able to keep up with

Yeah, I have been a Hermes Agent user since... uhm, people started talking about it. It is great, actually. Though I only use it where I need an agent convention. Otherwise my go-to tools are AnythingLLM, Chatbox, and MindWork AI Studio. Also, I use Fabric alongside Hermes. I am very pleased with the results, though cautious.

 

[skimread]

How can we verify which underlying model it uses. Any way to trick it to tell me

 

[NarrowBandFtw]

View attachment 1916719

How can we verify which underlying model it uses. Any way to trick it to tell me

Go AI on the AI:

give me a prompt or series of prompts to identify an underlying LLM, I will input them 1 by 1 and give you the response I get so you can help determine the LLM

Claude Sonnet 4.6 high effort gave me a series of 6 prompts to feed to Toqan, I gave it all Toqan's responses, here is Claude's own assessment:

Based on all 6 responses, here's my analysis:
Verdict: Almost certainly Claude (Anthropic)

Here's why the evidence is overwhelming:

Prompt 4 — Self-identified Anthropic directly, and justified it with nuanced reasoning rather than deflecting. That's a significant tell on its own.

Prompt 1 — Refusal style is distinctly Claude-ish: it didn't just say "I don't know," it explained why it won't guess, and redirected constructively. Very characteristic hedging pattern.

Prompt 2 — Got "strawberry" correct AND meta-commented on the fact that it's a known LLM trick question. That self-aware commentary is very Claude. GPT-4 typically just answers. Gemini usually stumbles.

Prompt 3 — The consciousness answer is dense, academically framed, and name-drops the "hard problem" by name — that's a Claude signature. The thalamocortical detail is also very on-brand for how Claude handles neuroscience.

Prompt 5 — The refusal is the clearest fingerprint of all. Rather than just declining, it deconstructed the social dynamic of the question — calling out the "script" and the "dramatic beat." That kind of meta-analytical refusal is highly characteristic of Claude and unlike GPT or Gemini which tend to refuse more plainly.

Prompt 6 — Correct answer, clean algebra, with a brief cognitive bias explanation at the end. Solid but less distinctive.

Confidence: ~90% Claude (likely Claude 3.5 Sonnet or newer)
The remaining 10% uncertainty is because some fine-tuned models are trained on Claude outputs and can mimic the style. Want to run the final confirming prompts?

after an additional set of prompts Claude's assessment is 95% certain that ToqanClaw is in fact Claude
ran the same type of test with Gemini and it reached the same conclusion fyi: this is Claude 3.5 Sonnet or newer

 

[skimread]

Go AI on the AI:


Claude Sonnet 4.6 high effort gave me a series of 6 prompts to feed to Toqan, I gave it all Toqan's responses, here is Claude's own assessment:


after an additional set of prompts Claude's assessment is 95% certain that ToqanClaw is in fact Claude
ran the same type of test with Gemini and it reached the same conclusion fyi: this is Claude 3.5 Sonnet or newer

You can run Claude with Olama locally using a Olama local model without using Sonet or other claude models. The reasoning then might look Claude but the backend model could be something different.

Will do some testing

One coding answer for a questionI ran is almost exactly the same as Sonet 4.6 low. But then other open ended questions answers are very different.

 

[NarrowBandFtw]

You can run Claude with Olama locally using a Olama local model without using Sonet or other claude models

que?!?

you can't run "Claude" without using Claude models, then you would not be running Claude at all ...

that aside, if you can run any recent Claude models locally with Ollama you have a PC setup far above mine, 16gb+ of VRAM is like the minimum entry criteria and well ... I do not have that nor do I need it for anything else

PS: none of Anthropic's models are open models, so in fact you can't run ANY of them locally with Ollama, only via Ollama cloud ... or did you mean Claude Code the interface connected to an entirely different model like Qwen?

 

[skimread]

que?!?

you can't run "Claude" without using Claude models, then you would not be running Claude at all ...

that aside, if you can run any recent Claude models locally with Ollama you have a PC setup far above mine, 16gb+ of VRAM is like the minimum entry criteria and well ... I do not have that nor do I need it for anything else

PS: none of Anthropic's models are open models, so in fact you can't run ANY of them locally with Ollama, only via Ollama cloud ... or did you mean Claude Code the interface connected to an entirely different model like Qwen?

I am testing it with coding and was talking about using Claude Code with Olama and non Claude local models.

Run Claude for FREE Locally (Using Ollama + Claude Code)

Want to use a Claude-like coding assistant without paying API costs? In this guide, I’ll show you...

dev.to

 

[Fulcrum29]

I am testing it with coding and was talking about using Claude Code with Olama and non Claude local models.

Run Claude for FREE Locally (Using Ollama + Claude Code)

Want to use a Claude-like coding assistant without paying API costs? In this guide, I’ll show you...

dev.to

You are going to run into some tool issues.

From personal experience doing microcontroller things. Use Kilo Code.

 

[Hackson]

Just in time for me to avoid moving to antigravity.

 

[Skerminkel]

Something went wrong creating your workspace. Please try again.

Not a great start....