DGX Spark benchmarks: which local model wins in 2026

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. For the full spec sheet and a photo, see the DGX Spark hardware page.

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, and the DGX Spark hardware page for the spec sheet. This page is the model-side companion: who wins when you actually run them.

The best open-weight models that fit the Spark

The benchmark sections below cover specific runs that change as suites add models. For a current, catalog-level answer to “what should I actually run on this box,” here are the strongest open-weight models with a variant that fits the Spark’s 128GB of unified memory. A model “fits” when its smallest comfortable variant (recommended memory, from the model catalog) stays well under 128GB, leaving headroom for context and the OS. Scores are Tokenstead’s quality ratings (coding / reasoning / general / tool-calling, 0-100), not this guide’s benchmark runs.

The newest releases (August 2026)

These are the open-weight models that shipped in the last month and fit the Spark - the ones to try first if you want the current state of the art on this box.

• Qwen3.8-27B - the pick. The newest and strongest all-round fit on the Spark: 88 coding / 89 reasoning / 94 tool-calling, 262K context, running at just ~21GB (Unsloth UD-Q4_K_XL). It combines near-flagship quality with a small enough footprint to leave the box plenty of headroom for long context, and Unsloth’s Dynamic 3.0 quants make it fast on this hardware. If you only install one model, start here. Released 2026-08-05. See Qwen3.8-27B.
• Pokee-Isaac 28B - 90 / 88 / 88 with a 95 tool-calling score and a 10M-token context, ~22GB (Q4_K_M). The highest quality-per-gigabyte of the current field, and a 10M context no other fit touches. If your workload needs huge context or the absolute best agentic tool-calling, this is the alternative to Qwen3.8-27B. Released 2026-08-04. See Pokee-Isaac 28B.
• Nemotron 3.5 Lightning - 78 / 76 / 74 with 80 tool-calling and a 1M-token context, ~40GB (Q4_K_M). The long-context specialist for agent workloads that need the whole repository in window. Released 2026-08-11. See Nemotron 3.5 Lightning.
• Muse Glimmer 30B - fits at ~36GB (Q4_K_M), with day-zero Ollama support. The newest of the group (2026-08-10); not yet independently benchmarked on the Spark, so treat it as a promising new entry rather than a proven pick. See Muse Glimmer 30B.

The proven all-rounders (earlier releases, still excellent)

These are the older models that remain the most battle-tested fits on the Spark, backed by the benchmark runs later in this guide.

• Qwen3.6 35B A3B - 90 coding / 91 reasoning / 95 tool-calling, 262K context, ~22GB (Q4_K_M). The MoE sweet spot: a small active-expert count keeps decode fast, and it is the model the NVFP4 community runs hit 200+ tok/s on. See Qwen3.6 35B A3B.
• Gemma 4 31B - 86 / 89 / 85, 256K context, ~22GB (Q4_K_M, QAT, or NVFP4). The best dense all-rounder that still fits comfortably; skip the FP16 path (10s+ TTFT) for interactive use. See Gemma 4 31B.
• Gemma 4 26B A4B - 85 / 88 / 84, 256K context, ~18GB (Q4_K_M). A touch less raw quality than the 31B but the smallest comfortable fit in the strong field, and it cleared 17/17 with 50+ tok/s in Exxact’s agent suite. See Gemma 4 26B A4B.
• Qwen3 235B A22B - 88 / 90 / 85, fits only at the aggressive Q2_K (~80GB). The largest model that still fits, and the pick for hard reasoning when you accept the quantization accuracy tradeoff. See Qwen3 235B A22B.

The takeaway: for most people, Qwen3.8-27B is the model to install first - the strongest all-round fit, newest, and cheap on memory. Reach for Pokee-Isaac 28B if you need 10M-token context or maximum tool-calling, and Nemotron 3.5 Lightning if you need a 1M-token window. The established MoE sweet spot (Qwen 3.6 35B, Gemma 4 26B) remains the proven fallback. The 235B MoE is the ceiling - it fits, but only at the aggressive quant, so weigh the accuracy cost before committing to it.

BridgeBench: the overall leaderboard

BridgeBench 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.

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.