TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Strata GitHub project says its software can run the 125-billion-parameter Qwen3.8-Flash-Next model on supported consumer PCs, using compressed model files and system RAM. Its published benchmark lists up to 94 tokens per second on an RTX 5070, not 100T/s on an RTX 4090; that specific headline claim is not substantiated by the supplied report.
The open-source Strata project says its software can run the 125-billion-parameter Qwen3.8-Flash-Next model on supported consumer PCs, but the report supplied for this story does not substantiate the claim of 100 tokens per second on an RTX 4090. Its published tests instead list a maximum of 94 tokens per second on an RTX 5070, with results varying by model compression and hardware.
Strata’s report describes a local setup for Windows or Linux PCs with supported Nvidia or AMD graphics cards. It says the model can chat, write code and process images, and that prompts and responses remain on the user’s computer. These are project statements; the source material does not include an independent audit of privacy or performance.
The listed Nvidia test used an RTX 5070 with 12 GB of VRAM, a Ryzen 5 7600 processor and 64 GB of system RAM. Strata reports answer-generation rates from 53 to 94 tokens per second across the listed model sizes. The 94-token result used Q2_0 compression and engine version 0.1.36; the other Nvidia results used version 0.1.26. The project says its tests used 4,000-token answers and 32,000-token prompts.
On an AMD system with an RX 9070 XT and 47 GB of RAM, the reported rates range from 44 to 60 tokens per second. Strata says the software needs at least 12 GB of graphics memory, 32 GB of system RAM and about 80 GB of free disk space. Its instructions describe a roughly 70 GB model download and say startup can temporarily slow the computer while 35–55 GB is loaded into RAM.
Large Models on Personal PCs
Running a model described as having 125 billion parameters without sending prompts to a hosted service could give users more control over where their data is processed. It could also reduce reliance on paid cloud access for people whose computers can handle the memory and setup requirements. Those benefits depend on Strata’s local-operation claims and the user’s own configuration.
The benchmark figures also show why the headline speed needs careful qualification. Strata’s own table reports different rates depending on compression, graphics card and software engine; it does not provide a 4090 result in the supplied material. Token-generation speed is only one part of performance, and the report’s prompt-processing figures use a particular 32,000-token test rather than describing every workload.
Products Worth Considering

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz

LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor

Samsung Galaxy Tab A11+ 8GB RAM, 256GB Storage, Gray
As an affiliate, we earn on qualifying purchases.
What Strata’s Benchmarks Cover
The report describes several compressed versions of the model. In general, its guidance says smaller compressed sizes run faster, while larger sizes can retain more of the model’s capability but require more memory. Strata recommends different options based on available system RAM; for example, it recommends a coding variant for a 32 GB system and lists several options for 64 GB.
Strata identifies its coding variant as having half of the model’s experts removed. It says the variant achieved 91% of the full model’s SWE-bench Verified score, as measured by the variant’s authors. That is an attributed result, not an independent comparison provided in the report. The project also cautions that the coding variant is weaker outside coding, including for Chinese and other CJK text.
The project says graphics cards with more VRAM may be faster and estimates that an RTX 3090 with 24 GB could generate about 100–140 tokens per second. That is a project estimate, not a measured RTX 4090 result in the supplied tables. The report also says multiple GPUs can share a model, while some older-card support is experimental and community-tested.
“A token is about ¾ of a word.”
— Strata project, GitHub report
The RTX 4090 Speed Claim
The source material does not show an RTX 4090 test, nor does it identify what “100T/s” means in the headline. The report uses tokens per second, but supplies no matching RTX 4090 configuration, model size, engine version or measurement for that rate. It is therefore unclear whether the headline refers to generation speed, another measure, or a separate test not included in the report.
The benchmark tables also do not provide independent replication, detailed measurement methodology or error ranges. The project notes that results depend on model size and hardware, but the supplied material does not establish how performance would vary across different 4090 systems or longer conversations. Its statement that data stays local is likewise a project claim; no external privacy review is included.
Evidence Needed for 4090 Results
To verify the specific speed claim, the project or another tester would need to publish a reproducible RTX 4090 benchmark naming the model variant, compression level, engine version, system RAM, prompt and output lengths, and measurement method. Independent tests across more than one system would help show whether the result is repeatable.
Until those details are available, the figures readers can attribute to Strata are its published RTX 5070 and RX 9070 XT results, alongside its estimate for an RTX 3090. The supplied source provides no confirmed follow-up test date or independent assessment of the software’s claims.
Key Questions
Does the report confirm 100 tokens per second on an RTX 4090?
No. The supplied benchmark tables contain no RTX 4090 result. Strata reports up to 94 tokens per second on an RTX 5070 using Q2_0, and separately estimates 100–140 tokens per second for an RTX 3090 with 24 GB of VRAM.
What speed did Strata report on an RTX 5070?
Strata lists 53–94 tokens per second for answer generation across the Nvidia model sizes shown. The 94-token result used Q2_0 and engine version 0.1.36; the other listed results used version 0.1.26.
What hardware does Strata say is required?
The project lists a supported Nvidia or AMD graphics card with at least 12 GB of VRAM, at least 32 GB of system RAM and about 80 GB of free disk space. The model download is listed at roughly 70 GB.
Does Strata say prompts are sent to a cloud service?
No. Strata says processing happens locally and that nothing leaves the PC. That is the project’s claim; the supplied report does not include an independent security or privacy audit.
Source: hn
