TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

XDA Developers tested Qwythos 9B v2, a free model from Empero AI, on an 8GB RTX 3070 and reported speeds of about 20–22 tokens per second using a quantized build. Empero says it trained the model on Claude conversations, but the source does not establish how that data was obtained or independently verify the claim. The test suggests Qwythos can handle some offline explanation and reasoning tasks, while current information requires web access and coding and vision remain limitations.

Qwythos 9B v2, a free model from Empero AI that the company says was trained on Claude conversations, ran on an 8GB Nvidia RTX 3070 in an XDA Developers test published October 24. The result suggests some users can try a local reasoning model on consumer hardware, but it does not verify Empero’s training-data claims or show that Qwythos can replace a Claude subscription across tasks.

XDA’s reviewer downloaded a Q4_K_M quantized version through LM Studio. The model file was just under 7GB, leaving room for additional graphics memory use by the context and other system demands. On the RTX 3070, the reviewer reported generation speeds of about 20–22 tokens per second, with many answers beginning in under a second. Those figures describe one test setup, not a performance guarantee for other GPUs or configurations.

Empero says Qwythos is based on Qwen3.5 9B and was trained on more than 500 million tokens of Claude Mythos and Claude Fable conversations. The article says the step-by-step reasoning examples were produced by Empero’s own tool, rethink. It also reports that Empero has not explained how it obtained the Claude material. The training description is therefore a company claim, not an independently verified account of the data or its provenance.

In the reviewer’s informal tasks, Qwythos solved a travel-time puzzle with thinking turned off, though it missed a departure-time detail in the prompt. With thinking enabled, it gave what the reviewer described as an accurate explanation of why coffee can lead to a crash, after taking 28 seconds to respond. It acknowledged lacking internet access when asked for the latest LM Studio version, and returned a correct answer after the reviewer connected Brave Search through an MCP tool. These tests were not presented as a controlled benchmark.

What an 8GB GPU Can Handle

The test addresses a practical barrier for people interested in running language models locally: graphics memory. A quantized 9-billion-parameter model that fits on an 8GB card gives users with older or mid-range hardware another option to explore, without relying on a cloud service for every prompt. The reported speed makes basic chat and explanation tasks appear usable on the tested machine.

That does not mean local access brings the same service as Claude. Qwythos runs without built-in access to live information, and the reviewer needed a separate search connection to answer a current-version question. Empero also acknowledges that coding performance is weaker than the underlying Qwen model and that Qwythos can state incorrect facts confidently without web search. The local model may suit some workflows, but the test does not establish broad parity with Claude or prove that it can replace subscription features.

How Qwythos Was Built

Qwythos v2 follows an earlier release. According to the source article, the revised version addresses a tendency in the original to repeat itself. It is described as a reasoning model, meaning it can generate an intermediate reasoning block before producing an answer; the model runner can allow users to switch that behavior on or off.

The underlying Qwen3.5 9B model already supports images and long documents, according to the article. Empero says it added training on Claude conversations, while its rethink tool generated the step-by-step reasoning material. The article describes Claude Mythos as a model available only to a limited group of vetted partners and says Fable 5 and Mythos 5 refer to the same Anthropic model. These details frame Empero’s account, but do not independently confirm the data source or training process.

The reviewer also reported that the model could be run with a 14,000-token context on the RTX 3070, depending on settings and how many model layers were placed on the GPU. The article cautions that a one-million-token context figure associated with the model is intended for server hardware, not the tested 8GB card.

“The company says Qwythos was trained on more than 500 million tokens of Claude Mythos and Claude Fable conversations.”

— Empero AI, as described in the XDA Developers report

Training Data and Test Limits

The largest unresolved issue is where the Claude conversations came from. The source report says Empero has not disclosed how it obtained the Mythos and Fable material. It does not establish whether Anthropic supplied or authorized the data, or whether the company’s description has been independently checked. No conclusion about the data’s provenance or permissions can be drawn from the report alone.

The performance evidence is also limited. XDA describes hands-on prompts rather than a broad, independently reproduced evaluation, and reports that Empero’s own seven-question web-search test was answered correctly without providing enough detail to judge its design. The article says vision was the weakest area but ends before giving the full test results. It is not clear how Qwythos performs across a representative set of image tasks, or how results compare under matched settings with Claude and other models.

What Users Should Verify

Readers considering Qwythos can check Empero’s model listing and documentation for the current v2 files, recommended settings, and any later explanation of its training data. Anyone testing it should distinguish offline answers from results produced with web search enabled, and should independently verify factual output, especially for medical, cybersecurity, or other consequential use.

Further independent evaluations could clarify how Qwythos compares with its Qwen base model on reasoning, coding, vision, and tool use. Until then, the reported RTX 3070 result is a useful indication that a quantized build can run on an 8GB card, not evidence of broad performance or a substitute for Claude’s hosted tools and services.

Key Questions

Can Qwythos 9B run on an 8GB graphics card?

In the XDA test, Qwythos 9B v2 ran on an 8GB RTX 3070 using a Q4_K_M quantized model file just under 7GB. Memory available for context and other processes depends on the setup.

How fast was it in the reported test?

The reviewer reported about 20–22 tokens per second on the RTX 3070, with many responses starting in under a second. These are results from one machine and configuration.

Was Qwythos trained on Claude data?

Empero AI says it was, citing more than 500 million tokens of Claude Mythos and Claude Fable conversations. The report says Empero has not explained how it obtained that material, and the claim is not independently verified in the source.

Can Qwythos provide up-to-date answers offline?

Not from the local model alone. In the reported test, it acknowledged it had no internet access; it answered a current-information question after the reviewer connected Brave Search through an MCP tool.

Does Qwythos replace a Claude subscription?

The report does not show that it does. Qwythos handled some explanation and reasoning prompts, but it needs added tools for current information, and the article notes limitations in coding and vision. The test did not compare the full services or subscription features.

Source: XDA Developers