TL;DR
The open-source Strata project says it can run the 125-billion-parameter Qwen3.8-Flash-Next model locally on supported consumer PCs, including systems with GPUs offering at least 12 GB of VRAM. Its published RTX 5070 results show up to 94 tokens per second for answer generation; the source does not substantiate the headline’s claim of 100 trillion tokens per second or provide an RTX 4090 benchmark.
The open-source Strata project says users can run the 125-billion-parameter Qwen3.8-Flash-Next model on supported consumer PCs, with published tests on an RTX 5070 and an AMD Radeon RX 9070 XT. The project’s figures show answer generation up to 94 tokens per second on the RTX 5070, but its supplied report does not document an RTX 4090 test or support the advertised claim of “100T/s.”
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Strata presents itself as a local launcher and runtime for the model, intended for Windows or Linux computers with supported NVIDIA or AMD graphics cards. Its requirements list at least 12 GB of GPU memory, 32 GB or more of system RAM, and about 80 GB of free disk space. The project says processing can take place on the user’s PC, rather than sending prompts to a remote service.
The report lists two test systems: an RTX 5070 with 12 GB VRAM, a Ryzen 5 7600 and 64 GB of RAM; and an RX 9070 XT with 16 GB VRAM, a Ryzen 9 3900X and 47 GB of RAM. On the NVIDIA system, the listed answer-generation rates range from 53 to 94 tokens per second across several compressed model sizes. The AMD results range from 44 to 60 tokens per second.
Those numbers measure different stages: “writes answers” refers to output generation, while “reads your prompt” refers to processing a 32,000-token prompt. The NVIDIA table uses engine 0.1.36 for its Q2_0 result and 0.1.26 for the other listed rows, with 4,000-token answers and 32,000-token prompts. Strata directs readers to its repository’s DETAILS.md for full tables.
Local AI on Gaming PCs
The reported results point to a practical shift: a model described as requiring server-class resources may be usable on a high-end home computer when compressed into smaller formats and paired with enough system memory. For users, local operation can mean keeping prompts and files on their own machines, though the project’s privacy description is a claim about its software setup, not an independent audit.
The output rates also help set expectations. At 53 to 94 tokens per second on the tested NVIDIA configurations, text may appear quickly for ordinary interactions, but speed depends on model size, hardware, prompt length and software version. The figures do not establish that every consumer PC can run the model at those rates, or that the compressed versions match the full model’s capabilities.
Products Worth Considering
Samsung Odyssey G9 QD-OLED 49 Dual QHD Curved Gaming Monitor
LG UltraGear 32-inch Curved QHD Gaming Monitor (2560×1440, 180Hz, 1ms, 1000R, HDR10)
Samsung 32 Odyssey G5 (G51F) QHD 180Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What Strata’s Benchmarks Cover
Strata offers several compressed formats, trading model size and speed against quality. Its guidance recommends different options depending on available RAM: for example, the project suggests a coding-focused variant for 32 GB systems and larger choices for machines with 64 GB or more. The repository says the Coder variant removes half of the model’s experts and cites its authors’ measurement of 91% of the full model’s SWE-bench Verified score; it also warns that Coder is weaker outside coding, including some Chinese and other CJK text.
The project’s published hardware list includes selected NVIDIA RTX 20, 30, 40 and 50 series cards and a range of AMD Radeon models. Strata says systems can combine two or three graphics cards, and it notes that startup may temporarily slow or make a PC unresponsive as 35–55 GB of data is loaded into RAM. The listed tests are project-reported benchmarks, not an independent comparison across consumer hardware.
“A token is about ¾ of a word.”
— Strata project, GitHub repository
The RTX 4090 Claim Is Unverified
The supplied source does not contain an RTX 4090 benchmark, and its measured results are stated in tokens per second, not trillions of tokens per second. The “100T/s” wording in the prompt is not substantiated by the repository material. Strata’s closest related figure is an estimate of 100–140 tokens per second for an RTX 3090 with 24 GB of VRAM, which should not be treated as a measured RTX 4090 result.
It is also unclear from the supplied material how the reported tests were independently verified, how performance varies across longer conversations or different settings, and how closely each compressed model format performs to the uncompressed model. The repository links to community results and detailed tables, but those do not establish a universal speed or quality level.
More Hardware Results Needed
The next useful evidence would be a reproducible RTX 4090 test that identifies the model format, Strata engine version, prompt and output lengths, and whether the result measures prompt processing or answer generation. Comparisons across multiple runs and model settings would help readers judge whether the performance generalizes beyond the project’s two listed test PCs.
Strata’s repository remains the place to check for updated installers, benchmark tables and community hardware reports. Until an RTX 4090 result is published with those details, the confirmed takeaway is narrower: the project reports local operation on supported gaming PCs, with its current listed RTX 5070 results topping out at 94 tokens per second.
Key Questions
Does the source confirm that Qwen3.8-Flash-Next runs at 100 trillion tokens per second?
No. The supplied Strata report gives results in tokens per second, with a listed maximum of 94 tokens per second on an RTX 5070. It does not support a rate of 100 trillion tokens per second.
Is there an RTX 4090 benchmark in the report?
No. The two systems with published results use an RTX 5070 and an AMD RX 9070 XT. The report does not include a measured RTX 4090 result.
What hardware does Strata list as a minimum?
Strata lists a supported NVIDIA or AMD graphics card with at least 12 GB of VRAM, 32 GB or more of system RAM, and about 80 GB of available disk space. Exact compatibility depends on the card and setup.
What does “local” mean in the project’s description?
Strata says the model runs on the user’s PC and that prompts do not leave it. That is the project’s description; the supplied material does not include an independent review of its privacy or network behavior.
Source: hn
