DGX Spark benchmarks: which local model wins in 2026

Independent benchmarks of NVIDIA's DGX Spark desktop through mid-2026 show Qwen 3.5 27B as the most consistent all-rounder on the easy Ollama path, while GPT-OSS 120B pushes nearly 4x the throughput but loses on reasoning and code generation. On the optimized NVFP4 path, Qwen 3.6 35B A3B hits 200+ tok/s with perfect tool-calling. The 273 GB/s bandwidth ceiling rewards small-active-parameter MoE and aggressive quantization for speed, while quality edge sits with denser models at full precision.

What this is The DGX Spark is NVIDIA’s GB10 Grace Blackwell desktop - 128GB of unified memory in a 1.2kg box - and the first question every new owner asks is “which model do I actually run on it?” The marketing shows capacity. Independent benchmarks show what that capacity buys you in real tok/s, latency, and task pass-rate. This guide collects the independent DGX Spark benchmark runs published through mid-2026 - BridgeBench, Exxact’s OpenClaw agent suite, and several community runs - so you can pick a model on numbers someone else measured, not a spec sheet. The short version up front: Qwen 3.5 27B is the most consistent all-rounder on the easy Ollama path; GPT-OSS 120B pushes nearly 4x the throughput but loses on reasoning and code generation; and on the optimized NVFP4 path, Qwen 3.6 35B A3B hits 200+ tok/s with perfect tool-calling. The split between “fits and runs” and “fits and wins” is the whole story on this box. For the hardware story - what the Spark gives you out of the box, the 273 GB/s bandwidth ceiling, NVFP4 setup, and the Rubin/Rosa/Feynman roadmap - see self-host your AI: DGX Spark, RTX builds, and Mac /guides/self-host-ai-dgx-spark-rtx-mac . This page is the model-side companion: who wins when you actually run them. BridgeBench: the overall leaderboard BridgeBench https://www.bridgebench.ai/dgx-spark runs an open-source-model track measured on a local DGX Spark - “real throughput, real latency, no cloud overhead.” Their overall leaderboard snapshot April 9, 2026 covers four models: - 1 - Qwen 3.5 27B FP16 : 76.3% pass, 11.1 tok/s, 361ms TTFT. The all-rounder. Wins reasoning 95.0% , code generation 75.0%, tied with Mistral , and the hallucination category 40.0%, tied with Mistral . - 2 - GPT-OSS 120B FP8 : 74.0% pass, 41.9 tok/s, 498ms TTFT. The throughput winner by a mile - 3.8x Qwen’s tok/s - and the only model to top instruction-following 80.0% . Loses on reasoning 86.7% and code 70.0% . - 3 - Mistral Small 4 23.6B, Q4 K M : 69.0% pass, 4.7 tok/s, 2910ms TTFT. Ties Qwen on code and hallucination at a quarter the size, but slow. - 4 - Gemma 4 31B FP16 : 64.0% pass, 16.5 tok/s, 10153ms TTFT. A brutal ~10-second time-to-first-token on the FP16 path - the kind of latency that rules a model out for interactive use regardless of quality. See Gemma 4 31B /models/gemma-4-31b . The headline insight: throughput and quality point in opposite directions on the Spark. GPT-OSS 120B streams tokens 3.8x faster than Qwen 3.5 27B yet scores lower on overall pass-rate, reasoning, and code. The 273 GB/s bandwidth ceiling rewards small-active-parameter MoE and aggressive quantization for speed, while the quality edge sits with the denser 27B at full precision. The model you pick depends on whether you are serving an interactive UI latency and tok/s matter or grinding through a batch pass-rate matters, latency does not . A caveat on the field: BridgeBench’s leaderboard is four models. It is a real, measured data point, not a comprehensive survey. Treat it as one vote. Exxact OpenClaw: the agent-suite results Exxact’s benchmark https://www.exxactcorp.com/blog/benchmarks/benchmarking-local-ai-agents-on-nvidia-dgx-spark May 2026 ran nine models on a DGX Spark through a 17-test structured agent suite T1-T17 plus a multi-hop chain-depth probe, using Ollama and the OpenClaw agent runtime. Pass-rate is tests cleared; hop depth is how many reasoning hops the model sustained before breaking the chain. - nemotron-3-super 120B-A12B: 17/17, 6-hop depth, 16.4 tok/s. The strongest overall agent profile - perfect tool-calling with the deepest multi-hop chain. - qwen3.5 27B: 17/17, 5-hop, 10.4 tok/s. Clean and reliable, but slow. - nemotron-3-nano 30B: 15/17, 5-hop, 64.7 tok/s. The best runtime fit for OpenClaw - fast with context headroom to spare, at the cost of two missed tests. - qwen3.5 35B-A3B: 17/17, 4-hop, 48.2 tok/s. Clean, reliable, and fast - the MoE model’s small active expert count keeps decode quick. - gemma4 26B: 17/17, 4-hop, 52.7 tok/s. The speed/reliability surprise - perfect score at over 50 tok/s. See Gemma 4 26B A4B /models/gemma-4-26b-a4b . - nemotron-3-nano 4B: 16/17, 4-hop, 64.2 tok/s. Fast, one noisy miss. - qwen3.5 122B-A10B: 17/17, 3-hop, 20.1 tok/s. Clean after a T11 timeout fix, but shallow multi-hop. - gemma4 e4B: 17/17, 2-hop, 52.6 tok/s. Fast but shallow. - gemma4 31B: 17/17, 2-hop, 9.7 tok/s. Clean but slow and shallow. The agent story: perfect tool-calling is a quality gate, not a gradient. A model that sometimes emits invalid JSON is operationally equivalent to one that always fails, because the agent framework cannot recover. Six of nine models cleared 17/17 - the differentiator is hop depth how far the model reasons before the chain breaks and tok/s how fast it loops . Nemotron 3 Super’s 6-hop depth at 16 tok/s is a different profile from Gemma 4 26B’s 4-hop at 52 tok/s - pick the former for hard multi-step reasoning, the latter for tight interactive agent loops. Community runs: NVFP4 changes the speed story The BridgeBench and Exxact numbers above are mostly the easy Ollama/GGUF path. On the optimized NVFP4 path through TensorRT-LLM or the Atlas engine, the Spark’s throughput picture changes dramatically for MoE models. - Qwen 3.6 35B via Atlas + NVFP4 spark-arena, May 23 2026 : 218.85 tok/s, 100/100 on Tool-Eval-Bench across all five categories, 130,753-token context, 74ms TTFT - rank 2 overall on the spark-arena leaderboard. That is roughly 4-5x the Ollama-path throughput for the same model family, with perfect tool-calling. See Qwen3.6 35B A3B /models/qwen3-6-35b-a3b . - ai-muninn’s 8-model run found the best agent stack was a pair: qwen3-coder-next Q4 K M at 47 tok/s as the primary agent, plus qwen3-vl 30B at 19GB for vision - 71GB combined, leaving the Spark headroom. The same run confirmed gpt-oss 120B failed tool-calling invalid JSON despite being 3x larger than the 32B winner - the quality-gate point again. It also found Q4 K M versus Q8 0 quality was “nearly invisible” across seven task categories. - Toolery 0.4.1 NVIDIA developer forums is a 143-scenario deterministic tool-calling benchmark that treats DGX Spark topology - single, dual, triple, quad, octa node - as a first-class ranking axis, and re-tiers scenarios by measured pass-rate. It is the most granular tool-calling coverage published and the natural next step if you want to probe a specific model’s agent reliability. The pattern across every community run: MoE models with small active expert counts 3-12B active are the Spark’s sweet spot. They read only the active experts per token, so decode is fast despite a large total parameter count, and NVFP4 amplifies that advantage. Dense 27-31B models score well on quality but pay for it in tok/s. 120B MoE models GPT-OSS, Nemotron 3 Super are fast and capable but hit the tool-calling quality gate harder. How to pick Match the model to the workload, not the leaderboard rank: - Interactive chat or coding assistant latency-sensitive : Qwen 3.6 35B A3B on the NVFP4 path if you will do the container setup, or Gemma 4 26B / Qwen 3.5 35B A3B on Ollama for a fast day-one. Avoid FP16 dense 31B - the 10s TTFT is a non-starter for interactive use. - Batch agent workflows quality over latency : Nemotron 3 Super 120B-A12B for deep multi-hop, or Qwen 3.5 27B for reliable single-hop. The pass-rate matters here, tok/s does not. - Maximum throughput serving: GPT-OSS 120B FP8 - 41.9 tok/s on BridgeBench, far ahead of the dense field - as long as your workload tolerates its lower reasoning score. - Tool-calling agents: pick a model with a measured 17/17 or 100/100, not a near-miss. The quality gate is binary. Every number on this page is a snapshot from the cited source’s run on their Spark. Verify against the live leaderboard before buying hardware around a specific figure - benchmark suites add models and re-run constantly, and a rank today is a rank today, not a verdict. To see what your specific rig can run Spark or otherwise , use the rig finder /find ; to browse every model with memory and quant fits, see the model catalog /models . For the hardware side - bandwidth, NVFP4 setup, the ConnectX-7 multi-node scaling - see self-host your AI: DGX Spark, RTX builds, and Mac /guides/self-host-ai-dgx-spark-rtx-mac .