What is the best model to run on an ASUS Ascent GX10?

Qwen 3.6 35B-A3B NVFP4 is the pick. The unsloth NVFP4-Fast recipe was tested on a real ASUS Ascent GX10 and measured 91.5 tok/s single-stream, 370.8 tok/s at 16 concurrent requests, and 707.0 tok/s aggregate at 64 concurrent, running vLLM 0.25.0 ARM64 with FP8 KV cache, MTP-3 speculative decoding, and a 262K context window. For maximum coding quality instead of speed, run DeepSeek V4 Flash NVFP4 with a trimmed context. Nemotron 3 Nano 30B-A3B and Gemma 4 26B MoE are the stable alternatives.

Is the ASUS Ascent GX10 the same as the NVIDIA DGX Spark?

The silicon is identical. Both use the NVIDIA GB10 Grace Blackwell superchip with 128GB of unified LPDDR5X at 273 GB/s, a 20-core Arm CPU, ConnectX-7 200Gb networking, and DGX OS. The differences are storage, chassis, and price: the GX10 ships a 1TB PCIe 4.0 SSD at $2,999 launch MSRP, while the DGX Spark Founders Edition ships 4TB and launched at $3,999 before rising to $4,699 in the early-2026 memory price spike. ASUS also sells a 4TB PCIe 5.0 SKU. Because the chip is the same, DGX Spark model benchmarks apply directly to the GX10.

Can you cluster ASUS Ascent GX10 units like DGX Sparks?

Yes. Each GX10 has ConnectX-7 200Gb QSFP networking. Two units connect back-to-back with a single QSFP cable and run tensor-parallel as one 256GB machine. ASUS states that four or more units can be clustered through a network switch. One caveat: early units shipped with SoC firmware that negotiated the second ConnectX-7 slot at PCIe Gen5 x2 instead of x4, cutting a Spark-to-GX10 link from about 220 Gbps to about 163 Gbps. BIOS GX10DGX.0104.2026.0326.1657 fixed it, so update firmware before you benchmark a cluster.

Is 1TB of storage enough for local LLMs on the GX10?

It is tight but workable if you are disciplined. A single NVFP4 checkpoint in the 128GB memory class occupies roughly 15-60GB on disk, so 1TB holds a working set of maybe eight to fifteen models alongside DGX OS, container images, and datasets. vLLM ARM64 images and HuggingFace caches consume space fast. If you plan to fine-tune, keep multiple checkpoint families, or store datasets locally, buy the 4TB SKU or add external NVMe over USB4 — the $1,000 saving disappears quickly when you are deleting models to make room.

← All guides

Hardware August 5, 2026

Best Models to Run on the ASUS Ascent GX10 (August 2026)

The ASUS Ascent GX10 runs the same NVIDIA GB10 Grace Blackwell superchip as the DGX Spark — 128GB of unified LPDDR5X at 273 GB/s, 1 PFLOP of FP4 — for $1,000 less. You give up storage, not compute. So every model recipe published for the Spark runs here unchanged, and the headline pick is the same: Qwen 3.6 35B-A3B NVFP4, measured at 91.5 tok/s single-stream on an actual GX10. This guide covers what the price gap actually costs you, which models to run at 128GB, and where the GX10's cheaper build shows up when you cluster.

Building a GB10 OpenClaw rig?

See our AI training options. We'll plan a GX10 or Spark-cluster OpenClaw setup that turns your GB10 boxes into a private AI server for your team.

Prefer the NVIDIA-branded box with 4TB storage? The reference machine is the NVIDIA DGX Spark 128GB.

Bottom Line (August 2026)

Best overall pick — Qwen 3.6 35B-A3B NVFP4. Measured on a real GX10 at 91.5 tok/s single-stream and 707 tok/s aggregate at 64 concurrent, with a 262K context window.
Best coding model that fits — DeepSeek V4 Flash NVFP4 at a reduced context. Same 128GB budget as the Spark, same result.
The GX10 trade — you save about $1,000 and lose 3TB of storage. Compute is identical. Nothing else about model selection changes.
Clustering works — ConnectX-7 200Gb, two units back-to-back as a 256GB machine, four or more through a switch. Update the BIOS first.

Every number below comes from published GB10 recipes and NVIDIA developer forum reports, mostly vLLM with NVFP4 quantization. Where a benchmark was run on a DGX Spark rather than a GX10, it still applies: the superchip, memory, and bandwidth are the same part.

The Hardware, Briefly

The GX10 is an NVIDIA GB10 Grace Blackwell superchip — a 20-core Arm CPU (10x Cortex-X925, 10x Cortex-A725) next to a Blackwell GPU with 48 SMs, rated at 1 PFLOP of FP4 — with 128GB of unified LPDDR5X at 273 GB/s. It runs DGX OS, ASUS says that is the only tested OS, and it draws 240W from a USB-C adapter in a 150 x 150 x 51 mm chassis.

That 273 GB/s is the constraint, exactly as on the Spark. Dense 70B-class models decode slowly. What wins on this box is MoE models with small active parameter counts (3B-15B active) quantized to NVFP4, Blackwell’s native 4-bit float format, which holds quality near FP8 at half the memory.

What the $1,000 actually buys

	ASUS Ascent GX10	NVIDIA DGX Spark FE
Chip	GB10 Grace Blackwell	GB10 Grace Blackwell
Memory	128GB LPDDR5X, 273 GB/s	128GB LPDDR5X, 273 GB/s
Storage	1TB PCIe 4.0 (4TB PCIe 5.0 SKU exists)	4TB PCIe 5.0
Networking	ConnectX-7 200Gb, 10GbE, Wi-Fi 7	ConnectX-7 200Gb, 10GbE, Wi-Fi 7
OS	DGX OS	DGX OS
Launch price	$2,999	$3,999, raised to $4,699 in early 2026

Storage is the whole difference. ServeTheHome put the 1TB GX10 at roughly $1,000 under NVIDIA’s 4TB model, which works out to about $333/TB for the storage you are giving up — cheap by NVMe standards, but you are also giving up PCIe 5.0 and Opal encryption on the drive. Street prices on both boxes have run above MSRP since the 2026 memory spike, so check current listings rather than launch figures.

Best Models on One GX10 (128GB)

Model	Quant	Context	Speed	Why
Qwen 3.6 35B-A3B	NVFP4	262K	91.5 tok/s (C1)	Verified on a real GX10; 370.8 tok/s at 16 concurrent
DeepSeek V4 Flash (284B/13B)	NVFP4	reduced	~55 typ / 78 peak	Best coding model that fits 128GB
Nemotron 3 Nano 30B-A3B	NVFP4	131K	~55-61 tok/s	NVIDIA’s own tune, most stable recipe
Gemma 4 26B MoE	NVFP4	128K	~52 tok/s	Best non-Qwen generalist at this size
Qwen 3.6 27B	NVFP4	256K	~33 tok/s	Leaves headroom for other workloads

Our pick: Qwen 3.6 35B-A3B NVFP4. The unsloth NVFP4-Fast build is the one recipe explicitly validated on ASUS hardware — the maintainer states it was tested on a real ASUS Ascent GX10 / NVIDIA GB10 with 128GB unified memory. The stack is vLLM 0.25.0 ARM64, FlashInfer B12X for the quantized MoE, FP8 KV cache, native MTP-3 speculative decoding, 262K context, with tool calling and image input working.

Read the concurrency numbers carefully, because they are the reason to buy a GB10 box over a Mac. Single-stream is 91.5 tok/s. At 16 concurrent requests the machine delivers 370.8 tok/s, and at 64 concurrent it delivers 707.0 tok/s aggregate. If you are serving a team or running multi-agent OpenClaw workloads, that scaling is the product.

Run DeepSeek V4 Flash NVFP4 instead when coding quality beats speed. It tops SWE-bench among open weights and squeezes into 128GB at NVFP4 with a trimmed context window.

One caution on NVFP4 maturity: as of the April 2026 forum consensus, NVFP4 kernels were still landing on GB10, and some builds underperformed their theoretical ~55-60 tok/s. FP8 remains the more predictable path if a recipe misbehaves — a benchmarked Qwen 3.6 35B-A3B FP8 vLLM serving config exists for exactly this reason.

If you use Ollama, switch to vLLM

A user running Llama 3.3 70B on a single GX10 measured vLLM against Ollama directly. vLLM served 82 tok/s generation against Ollama’s 47, 5,847 tok/s prompt processing against 1,268, and 243ms time-to-first-token against 1,412ms. Under 24 concurrent requests the gap widened. Ollama is fine for a first boot; it is not what you should serve from.

Clustering GX10 Units

Each GX10 carries ConnectX-7 200Gb QSFP networking. Two units connect back-to-back with one cable and run tensor-parallel as a single 256GB machine. ASUS states that four or more units cluster through a network switch. That unlocks the same tier ladder the Spark community has mapped:

2 units (256GB) — the sweet spot. DeepSeek V4 Flash DSpark at the full 1M-token context, roughly 42 tok/s at 1M and 44-63 tok/s at working contexts with MTP speculative decode. Omni alternatives: Inkling-Small and MiMo-V2.5 Omni, both at 1M context.
3 units (384GB) — GLM-5.2 NVFP4+AQLM at about 380K context, including the community vision build. Or DeepSeek V4 Flash on two units with the third free for ComfyUI, embeddings, and small agent models.
4 units (512GB) — unpruned GLM-5.2, 744B total / ~40B active, at up to 1M context and about 42 tok/s peak decode.

At $2,999 a unit, four GX10s cost roughly $12,000 against about $16,000 in Sparks for the same 512GB. That is the strongest argument for the ASUS box: the savings compound with every node.

Update the BIOS before you benchmark a cluster. Early GX10 units shipped with SoC firmware that negotiated the second ConnectX-7 slot at PCIe Gen5 x2 instead of x4. A Spark-to-Spark link ran about 220 Gbps while a Spark-to-GX10 link over the affected NIC ran about 163 Gbps, a 25% hit on any NCCL all-reduce pair touching that port. BIOS GX10DGX.0104.2026.0326.1657 restored Gen5 x4 on both slots. Flash first, then test.

Should You Buy the GX10 or the Spark?

Buy the GX10 if you are building a cluster, if 1TB of local storage covers your model working set, or if you simply want the cheapest route to GB10 compute. Four nodes saves you around $4,000.

Buy the DGX Spark if you fine-tune, hoard checkpoints, or want PCIe 5.0 storage with Opal encryption. Deleting models to free disk space is a worse daily experience than most people expect, and 4TB removes the problem entirely.

Buy neither if you will only ever run one box for single-user chat. A Mac Studio at 128GB has roughly twice the memory bandwidth (546 vs 273 GB/s), so dense models decode faster there. The GB10 platform wins on CUDA, on NVFP4 and vLLM tooling, on concurrent throughput, and on clustering — none of which a Mac can match.