What does Gabe Ortiz run locally, and on what hardware?

An AMD Ryzen 9 9900X with 128 GB of DDR5, an RTX 5090 32 GB and an RTX 3090 24 GB in the same box. On the RTX 5090, q27 (Quasar) runs Qwen3.8-27B at the 4-bit default tier (17.00 GB) with the DFlash2 block-diffusion drafter, about 4.1 accepted tokens per round on Claude Code traffic: ~220 tokens per second with speculative decoding (plain ceiling ~105 tokens per second). The RTX 3090 runs Whisper large v3. He has also been testing the IQ3_XXS quant of Qwen 3.8 Flash Next with Strata (https://github.com/Niko1221/Strata). Self-reported, 25 October 2026.

Where does he look for memory and tokens-per-second numbers?

Community benchmarks on the specific inference engine he wants to use. For the Mac, he has been looking at benchmarks for https://github.com/antirez/ds4, because it would probably be used for Deepseek.

Can I test open models in the browser for free before buying an RTX 5090 or RTX 3090?

Yes. You can test 10 open-weight models (including Gemma 4 31B, Gemma 4 26B A4B, Recursive Flint, and Nemotron 3.5 Super VL) side by side for free in your browser at https://localllmrig.com/ with no login and 0 GB of local VRAM, and check full Qwen3.8-27B memory requirements at https://localllmrig.com/can-i-run/qwen3.8-27b.

RTX 5090 32 GB + RTX 3090 24 GB Workstation Running q27 (Quasar) & Whisper Large v3

One AMD Ryzen 9 9900X workstation with 128 GB of DDR5, an RTX 5090 32 GB and an RTX 3090 24 GB in the same box. q27 (Quasar) is on the RTX 5090. Whisper large v3 is on the RTX 3090. A separate 8 GB RTX 3060 Ti runs Ternary Bonsai 2.

Gabe Ortiz (@signalnine), quoted by name with written permission. 25 Oct 2026. Self-reported. Not a localLLMrig measurement.

Test Gemma 4 31B, Recursive Flint and Nemotron 3.5 Super VL side by side in your browser, free.

gabeortiz.net. signalnine/q27#benchmarks (Quasar CUDA inference engine).

The rig

I also run a 24 GB RTX 3090 in the same box as the RTX 5090. It's an AMD Ryzen 9 9900X with 128 GB of DDR5. I run a transcription model on the RTX 3090: Whisper large v3. I've been playing with the IQ3_XXS quant of Qwen 3.8 Flash Next using Strata (https://github.com/Niko1221/Strata) a bit as well.

On the RTX 5090 he runs q27 (Quasar) with Qwen3.8-27B at the 4-bit default tier (17.00 GB) and the DFlash2 block-diffusion drafter. The drafter accepts about 4.1 tokens per round, measured on Claude Code traffic. Speed is ~220 tokens per second with speculative decoding (plain ceiling ~105 tokens per second).

The 8 GB RTX 3060 Ti

A separate machine, user-reported, and a different model from the RTX 5090 run. PrismML Ternary Bonsai 2 27B is a 1.6-bit T3_G128 pack (5 trits per byte, 6.06 GB) with no drafter. Decode is 42 tokens per second at 36.9K tokens of context, and prefill is 481 tokens per second. The closest 3060-class card on our list is a different board, the 12 GB one.

New model hardware requirements and price updates by email. We email a link to confirm the address. Nothing is sent until you click it. Privacy notice.

The next machine

Thinking about a Mac Studio M5 Ultra 256 GB as a home solution for running larger models.

Where he looks before buying

Community benchmarks on the specific inference engine I want to use. For the Mac, I've been looking at some of the benchmarks for https://github.com/antirez/ds4, because it would probably be used for Deepseek.

What would make a speed figure trustworthy

Hardware configuration and software config specifics: power limits set, RAM speed, disk speed, and so on.

What a hardware-fit site should state up front

Engine, spec-decode batch and quant are VERY important. Memory bandwidth. Context size limits. Chat vs agentic are very different workloads with different decode characteristics with regard to caching and speculative decode - specify which.

On our list: RTX 5090 32GB, RTX 3090 24GB, RTX 3060 12GB, Mac Studio M5 Ultra 256GB, Can I run Qwen3.8-27B?.

The numbers on this page were measured by the owner on that hardware. They are not a localLLMrig benchmark, and they do not change the fit scores in the calculator.

Questions

What does Gabe Ortiz run locally, and on what hardware?: An AMD Ryzen 9 9900X with 128 GB of DDR5, an RTX 5090 32 GB and an RTX 3090 24 GB in the same box. On the RTX 5090, q27 (Quasar) runs Qwen3.8-27B at the 4-bit default tier (17.00 GB) with the DFlash2 block-diffusion drafter, about 4.1 accepted tokens per round on Claude Code traffic: ~220 tokens per second with speculative decoding (plain ceiling ~105 tokens per second). The RTX 3090 runs Whisper large v3. He has also been testing the IQ3_XXS quant of Qwen 3.8 Flash Next with Strata (https://github.com/Niko1221/Strata). Self-reported, 25 October 2026.
How fast is Ternary Bonsai 2 on the 8 GB RTX 3060 Ti?: That is a separate machine and a different model, user-reported. PrismML Ternary Bonsai 2 27B is a 1.6-bit T3_G128 pack (5 trits per byte, 6.06 GB) with no drafter. Decode is 42 tokens per second at 36.9K tokens of context, and prefill is 481 tokens per second.
What machine is Gabe Ortiz weighing next?: Thinking about a Mac Studio M5 Ultra 256 GB as a home solution for running larger models.
Where does he look for memory and tokens-per-second numbers?: Community benchmarks on the specific inference engine he wants to use. For the Mac, he has been looking at benchmarks for https://github.com/antirez/ds4, because it would probably be used for Deepseek.
What would make a tokens-per-second figure trustworthy?: Hardware configuration and software config specifics: power limits set, RAM speed, disk speed, and so on.
What should a hardware-fit site state up front?: Engine, spec-decode batch and quant are VERY important. Memory bandwidth. Context size limits. Chat vs agentic are very different workloads with different decode characteristics with regard to caching and speculative decode - specify which.
Can I test open models in the browser for free before buying an RTX 5090 or RTX 3090?: Yes. You can test 10 open-weight models (including Gemma 4 31B, Gemma 4 26B A4B, Recursive Flint, and Nemotron 3.5 Super VL) side by side for free in your browser at https://localllmrig.com/ with no login and 0 GB of local VRAM, and check full Qwen3.8-27B memory requirements at https://localllmrig.com/can-i-run/qwen3.8-27b.

Find hardware · All builder interviews · Send us your llama-bench result