What informed the move from an RTX 4090 to the DGX Spark?

The priority was GPU-accessible memory capacity, so he could experiment with a wider range of models. That is why he moved from an RTX 4090 desktop to the Spark, despite its memory-bandwidth limitations. Linux ARM64 support has been challenging, and he says it seems to be improving. Upstream ds4, with kernel-level optimizations for this hardware, matters to what the machine can actually do.

What would make a tokens-per-second figure trustworthy?

He wants the experimental conditions transparent and reproducible: the exact model and quantization, engine version, workload, and hardware settings, rather than a single tokens/s number. In his own reporting he keeps the GPU clock range at 300-2200 MHz and runs at least three repetitions. He also wants prefill and decode reported separately, with the context length and the cache conditions.

What should a hardware-for-this-model page show besides a memory fit?

A GGUF file does not automatically run in llama.cpp or every other GGUF-capable engine. His releases can use custom architectures or quantization formats that need a specific runtime, so the file format alone does not say what will work. For each release he wants the supported inference engines and required versions, any extra files or setup, and which hardware configurations have actually been tested. He wants 'fits in memory' distinguished from 'supported and tested with this engine', with links to reproducible results.

May we quote Sunghoon Baek by name?

Yes. On 27 October 2026 he wrote: OK, please go ahead and publish the full interview with my five answers and the attribution and project links you described. The handle is @Baekpica. GitHub is https://github.com/Baekpica and Hugging Face is https://huggingface.co/Baekpica.

Can I test open models in the browser before buying a DGX Spark?

Yes. You can test open-weight models side by side in your browser at https://localllmrig.com/. Reflection Beam 1.0 is at https://localllmrig.com/can-i-run/beam-1.0 and Qwen3.8-Flash-Next is at https://localllmrig.com/can-i-run/qwen3.8-flash-next.

NVIDIA DGX Spark (128GB) local LLM: Beam-501B and Qwen3.8-Flash-Next tokens per second with ds4-dfm-rs (Sunghoon Baek, @Baekpica)

One Lenovo ThinkStation PGX, the GB10-based DGX Spark OEM system, with 128 GB of unified memory and a 4 TB SSD, bought earlier in 2026 for roughly KRW 6.3 million. He runs Beam-1.0-501B and his own Qwen3.8-Flash-Next quantizations on it with ds4-dfm-rs, a Rust-host continuation of ds4, and he moved from an RTX 4090 24 GB desktop for the memory capacity. When he reports local LLM speed in tokens per second he locks the GPU clock range to 300-2200 MHz, runs at least three repetitions, and separates prefill from decode with the context length and cache state. This page does not add a tokens-per-second figure he did not publish. Self-reported, community measurement, not ours.

Sunghoon Baek (@Baekpica), quoted by name with written permission. 27 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.

github.com/Baekpica. huggingface.co/Baekpica. ds4-dfm-rs. VQ collection for DGX Spark. MXQ collection for DGX Spark. Qwen3.8-Flash-Next Mixed-Quant SSD-PLE.

Beyond Beam-1.0-501B on a single DGX Spark, what else do you run locally today - models, quants, serving framework?

Here are my answers. A little context: I work professionally in AI, and my current projects specifically target a single DGX Spark.

Beyond Beam, I publish my own VQ and mixed-quantization releases for a single DGX Spark. I also maintain and use ds4-dfm-rs, a Rust-host continuation of ds4 (DwarfStar), to serve them:

VQ collection: https://huggingface.co/collections/Baekpica/ds4-vq-for-dgx-spark

MXQ collection: https://huggingface.co/collections/Baekpica/ds4-mxq-for-dgx-spark

ds4-dfm-rs: https://github.com/Baekpica/ds4-dfm-rs

One release I'd particularly like to highlight is Qwen3.8-Flash-Next Mixed-Quant SSD-PLE (https://huggingface.co/Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF). It is a useful example of pairing a hardware-specific quantization with the engine support needed to run it on one Spark.

What are you weighing for your next machine or upgrade, and what is the budget roughly?

I bought a Lenovo ThinkStation PGX, the GB10-based DGX Spark OEM system with a 4 TB SSD, for roughly KRW 6.3 million earlier in 2026. Recently, I've seen prices above KRW 10 million, so I'm not planning an upgrade for now. My near-term plan is to keep releasing quantized versions of newer models optimized for this machine.

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.

What informed your hardware choice?

My main priority was GPU-accessible memory capacity, because more capacity lets me experiment with a wider range of models. That is why I moved from an RTX 4090 desktop to the Spark, despite its memory-bandwidth limitations.

Software support matters too. Linux ARM64 has been challenging from the beginning, although it seems to be improving. Projects such as upstream ds4, with kernel-level optimizations targeting this hardware, are also important to what the machine can actually do.

What would make you trust a tokens/s figure you didn't measure yourself?

For me, the key is making the experimental conditions as transparent and reproducible as possible. I want to know the exact model and quantization, engine version, workload, and hardware settings, rather than just see a single tokens/s number.

In my own reporting, I keep the GPU clock-range setting consistent at 300–2200 MHz and run at least three repetitions. It also matters whether the number measures prefill or decode, and under what context length and cache conditions.

Anything you wish a 'which hardware for this model' site told you?

A clear, at-a-glance compatibility view would be especially useful. A GGUF file does not automatically mean it will run in llama.cpp or every other GGUF-capable engine. Releases such as mine can use custom architectures or quantization formats that require specific runtime support, so the file format alone does not tell users what will actually work.

For each release, I'd like to see the supported inference engines and required versions, any additional files or setup, and which hardware configurations have actually been tested. Clearly distinguishing 'fits in memory' from 'supported and tested with this engine' would be particularly helpful, alongside links to reproducible results. That would make it easier to understand what an existing machine can really run.

On our list: Can I run Reflection Beam 1.0?, Can I run Qwen3.8-Flash-Next?, NVIDIA DGX Spark 128GB, GeForce RTX 4090 24GB.

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

Beyond Beam-1.0-501B on a single DGX Spark, what else do you run locally today - models, quants, serving framework?: Beyond Beam, he publishes his own VQ and mixed-quantization releases for a single DGX Spark and serves them with ds4-dfm-rs, a Rust-host continuation of ds4 (DwarfStar). VQ collection: https://huggingface.co/collections/Baekpica/ds4-vq-for-dgx-spark. MXQ collection: https://huggingface.co/collections/Baekpica/ds4-mxq-for-dgx-spark. ds4-dfm-rs: https://github.com/Baekpica/ds4-dfm-rs. One release he highlights is Qwen3.8-Flash-Next Mixed-Quant SSD-PLE (https://huggingface.co/Baekpica/Qwen3.8-Flash-Next-Mixed-Quant-SSD-PLE-GGUF), a hardware-specific quantization with the engine support needed to run it on one Spark. His current projects target a single DGX Spark. Self-reported, 27 October 2026.
What are you weighing for your next machine or upgrade, and what is the budget roughly?: He bought a Lenovo ThinkStation PGX, the GB10-based DGX Spark OEM system with a 4 TB SSD, for roughly KRW 6.3 million earlier in 2026. He has recently seen prices above KRW 10 million, so he is not planning an upgrade for now. The near-term plan is to keep releasing quantized versions of newer models optimized for this machine.
What informed the move from an RTX 4090 to the DGX Spark?: The priority was GPU-accessible memory capacity, so he could experiment with a wider range of models. That is why he moved from an RTX 4090 desktop to the Spark, despite its memory-bandwidth limitations. Linux ARM64 support has been challenging, and he says it seems to be improving. Upstream ds4, with kernel-level optimizations for this hardware, matters to what the machine can actually do.
What would make a tokens-per-second figure trustworthy?: He wants the experimental conditions transparent and reproducible: the exact model and quantization, engine version, workload, and hardware settings, rather than a single tokens/s number. In his own reporting he keeps the GPU clock range at 300-2200 MHz and runs at least three repetitions. He also wants prefill and decode reported separately, with the context length and the cache conditions.
What should a hardware-for-this-model page show besides a memory fit?: A GGUF file does not automatically run in llama.cpp or every other GGUF-capable engine. His releases can use custom architectures or quantization formats that need a specific runtime, so the file format alone does not say what will work. For each release he wants the supported inference engines and required versions, any extra files or setup, and which hardware configurations have actually been tested. He wants 'fits in memory' distinguished from 'supported and tested with this engine', with links to reproducible results.
May we quote Sunghoon Baek by name?: Yes. On 27 October 2026 he wrote: OK, please go ahead and publish the full interview with my five answers and the attribution and project links you described. The handle is @Baekpica. GitHub is https://github.com/Baekpica and Hugging Face is https://huggingface.co/Baekpica.
Can I test open models in the browser before buying a DGX Spark?: Yes. You can test open-weight models side by side in your browser at https://localllmrig.com/. Reflection Beam 1.0 is at https://localllmrig.com/can-i-run/beam-1.0 and Qwen3.8-Flash-Next is at https://localllmrig.com/can-i-run/qwen3.8-flash-next.

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