Buying a local AI machine is easy. Choosing the right model for the number of machines you own is harder. That is the part that leaves hardware sitting idle. In September 2026, the DGX Spark community published a wave of deployment recipes and benchmarks. We can now answer that question with real numbers instead of marketing.

TL;DR

  • One DGX Spark (128 GB unified memory): Qwen3.8 Flash Next wins almost by default.
  • Two to four Sparks: GLM 5.3 Flash and DeepSeek v4.1 Flash share the base layer. Qwen3.8 Flash Next works as the worker.
  • Six Sparks and up: orchestrator, worker and security scanner — a small AI company on your desk.

What is DGX Spark, and why does it change the maths?

DGX Spark is NVIDIA's AI workstation. It runs on the Grace Blackwell GB10 chip. The headline feature is 128 GB of unified memory. The GPU and CPU share one pool of memory. Inference draws about 38W, according to NVIDIA's official documentation. A single unit lists for around $4,000.

The number matters less than the consequence. Several Sparks can link directly to each other with no network switch. They pool memory into 128, 256, 384 or 512 GB. That is how mixture-of-experts (MoE) models start running on a desk. These models carry hundreds of billions of parameters.

MoE is an architecture that activates only a small slice of its parameters. The rest stay idle for each token. The community moved fast. Within weeks, GitHub and the NVIDIA Developer forums held recipes for almost every cluster size. Here is the picture in September 2026.

One Spark: Qwen3.8 Flash Next is nearly the only answer

On a single Spark, most users run Qwen3.8 Flash Next. Alibaba's open model carries about 176B parameters. That breaks down as 125B core plus 51B n-gram. It activates only 6B per token, thanks to that MoE design.

The 6B active figure explains the fit. The smallest quantized build needs just 78 GB. That sits inside one Spark with room to spare. Community numbers from September 2026:

  • NVIDIA's official NVFP4 recipe on vLLM hits 64 tok/s peak for a single stream. It holds that speed on one, two and four machines.
  • Users running real coding work report about 43 tok/s on one machine (NVIDIA DGX Spark / GB10 forum).
  • Context support reaches 262K tokens, with multimodal input. Unsloth also reports strong agentic coding results.

One technical detail deserves respect. A community recipe served a 126 GiB checkpoint on a machine with only 121.63 GiB of real memory. It did that by tuning SGLang and leaning on the RAM-friendly design. Qwen3.8 Flash Next was built for unified memory. That is Spark's home ground.

The conclusion for this tier is short. If you own one machine, stop deliberating. Install Qwen3.8 Flash Next.

Two Sparks: GLM 5.3 Flash and DeepSeek v4.1 Flash step in

Pooled 256 GB opens the door to frontier-class models. Two names dominate at this level. Both ship under the MIT licence.

GLM 5.3 Flash comes from Z.ai. It carries 320B parameters with 18B active. Its FP8 checkpoint runs about 306 GiB. A community recipe runs it on two Sparks with vLLM TP=2. Tensor parallel splits the model across both machines. That setup logged 43.4 tok/s at 262K context. It accepts images and video, and its synthesis quality is strong.

DeepSeek v4.1 Flash shipped on 10 September 2026. It uses an encoder-decoder design with 552B parameters. The FP8 checkpoint runs about 510 GB. The earlier v4 Flash was the most popular pick on dual Sparks. It ran at 55–60 tok/s with 1M context. Version 4.1 is much heavier. Some users keep the older checkpoint, because it suits exactly that pair.

On the NVIDIA forums, head-to-head comparisons land on a modest verdict. The two models perform at nearly the same level in real work. One user ran both side by side through OpenWebUI. They found GLM slightly ahead because of vision. DeepSeek felt a little faster. Their advice was to keep a dedicated vision model either way.

The operating logic at two machines stays simple. One model, using both machines. Role splitting can wait.

Three and four Sparks: the orchestrator and worker split

At 384–512 GB, a pattern appears. Users stop running one giant model across the whole cluster. They split the roles instead.

The most common three-machine setup:

  • GLM 5.3 Flash on two machines as the orchestrator — it takes the task, plans, and hands out work.
  • Qwen3.8 Flash Next on one machine as the worker — it writes code, reads documents, and answers.

Four machines follow the same logic with more room:

  • GLM 5.3 Flash on two machines (orchestrator) plus Qwen3.8 Flash Next on two (worker pool), or
  • Run DeepSeek v4.1 Flash or GLM 5.3 Flash across all four machines. That gives you one model as strong as possible.

Why does splitting beat one huge model? Three reasons come up again and again in community reports:

  1. Lower latency. A small worker answers fast. The orchestrator wakes only to decide.
  2. No cascading failure. A worker that struggles does not take down the control plane.
  3. Independent upgrades. When a new Qwen ships, you swap the worker and keep the orchestrator.

This is the pattern agent teams increasingly prefer. A large model thinks, and small models act.

Six Sparks and up: a small AI company on your desk

From six machines (768 GB pooled and beyond), you can run three roles at once. Many users land on this setup:

  • GLM 5.3 Flash on two machines — orchestrator: plans, assigns, and reviews results.
  • Qwen3.8 Flash Next on two machines — worker: runs daily tasks at high speed.
  • DeepSeek v4 Flash on two machines — security scanner: reviews code and analyses threats. It also checks the output of the other two.

One subtle point rarely gets mentioned. The security role does not need a creative model. It needs discipline, low hallucination and careful reading. DeepSeek fits that job well. Structured reasoning and long context are its strengths. Qwen3.8 Flash Next handles fast repetitive work at 6B active and 43+ tok/s. That is the right call for performance per watt.

At this scale, total power draw stays in the low hundreds of watts. That is less than one datacenter GPU. Some companies refuse to send data to the cloud. For them, this answers a real question. Where does private AI actually run?

Insight: three things a spec sheet never tells you

One — checkpoint size is the binding constraint, not parameter count. GLM 5.3 Flash weighs roughly 306 GiB in FP8. DeepSeek v4.1 Flash runs about 510 GB. That decides how many machines a model needs. Speed only enters the conversation after that.

Two — community recipes matter as much as memory. Market Intelligence Research published a sizing guide in September 2026. It shows the densest recipe ecosystem sits at two and three machines. A middling model with a well-debugged recipe often beats a near-fit model that nobody has figured out yet. Before buying a fourth machine, check whether anyone has run your model on four.

Three — high-quality quantization is Spark's signature. Large unified memory lets NVFP4 and FP8 builds keep quality close to the original. That differs from forcing a large model into a 24 GB GPU with 2-bit quantization. The model visibly dulls.

Takeaway

If you are weighing a local AI investment right now, the rule is simple.

MachinesWhat to runWhy it works
1 SparkQwen3.8 Flash NextFits inside 128 GB and gets you working fastest
2 SparksGLM 5.3 Flash or DeepSeekFrontier quality, one model across both machines
3–4 SparksGLM as orchestrator + Qwen as workerLower latency and independent upgrades
6+ SparksAdd DeepSeek as security scannerThree roles, one small AI company

The specific next step: pick your machine count on the map above. Then find the matching recipe on GitHub. Names like qwen3.8-Flash-DGX and GLM-5.3-Flash-NVFP4-262K-2x-DGX-Spark are public. Run it for one week on real work before you buy more hardware. A local AI cluster is like hiring. You try before you sign a long contract.

If your team needs help designing local AI architecture, talk to 5ac.vn. We build and run systems like this on private infrastructure every day.

Sources