Best Open Source Models for NVIDIA GB10: Text, Image, and Video (August 2026)

The GB10 Architecture: Why Unified Memory Matters

The NVIDIA Grace Blackwell (GB10 / DGX Spark) architecture provides 128GB of unified LPDDR5X memory shared seamlessly between the ARM CPU cores and Blackwell GPU streaming multiprocessors. Unlike discrete PCIe GPUs that bottleneck on 32GB or 48GB VRAM boundaries, the GB10 can allocate up to 90GB of contiguous VRAM to a single inference process.

Here is our curated, tested lineup of the state-of-the-art open source models across Text, Image, and Video running locally on GB10.

---

1. Text & Agentic Reasoning Category

A. NVIDIA Nemotron 3.5 Lightning (August 2026) — The 1M-Context Execution Speedster

Released on August 11, 2026, Nemotron 3.5 Lightning is NVIDIA's purpose-built "execution layer" engine for high-volume agentic loops, sub-agent delegation, and tool calling:

• Architecture: Hybrid Mamba-2 State Space + Mixture-of-Experts (MoE) + Attention.
• Parameters: 30 Billion total parameters with only ~3 Billion active parameters per token.
• Context Window: 1,000,000 (1M) tokens native context.
• GB10 Optimization: In NVFP4 quantization (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark), it consumes only ~18GB VRAM and delivers lightning-fast token generation for repetitive tool loops without bloating KV cache memory.

B. Qwen 3.6 / 3.8 Series — The Deep Synthesis & Math Workhorse

For deep code synthesis, symbolic constraint solving, and competitive problem solving (such as the ARC Prize 2026):

• Model: nvidia/Qwen3.6-35B-A3B-NVFP4 / Qwen3.8-27B.
• Context: Up to 262,144 tokens.
• Performance on GB10: 4 concurrent batch sequences served via vLLM with FlashInfer and Marlin MoE kernels.

---

2. Image Generation & Visual Synthesis Category

A. FLUX.1 [dev] & [schnell] (Black Forest Labs)

The 12-billion parameter Rectified Flow Diffusion Transformer (DiT) remains the gold standard for photorealism, typography rendering, and complex prompt adherence:

• VRAM Footprint on GB10: In FP8/NVFP4 weights via ComfyUI on port 8188, FLUX.1 loads cleanly in ~24GB VRAM.
• Throughput: Generates 1024x1024 high-resolution images in 4 steps ([schnell]) or 20 steps ([dev]) in under 3.5 seconds.

B. Stable Diffusion 3.5 Medium (Stability AI)

A lightweight 2.5B parameter Multimodal Diffusion Transformer (MMDiT) capable of sub-second iteration on local hardware with zero memory contention.

---

3. Video Generation Category

A. CogVideoX-5B (THUDM)

An open-weights text-to-video diffusion transformer generating smooth 720p 6-second video sequences:

• Memory Requirement: ~16GB VRAM in FP8 with 3D VAE tile decoding.
• GB10 Fit: Easily fits within the unified memory pool, producing temporal coherent video without frame jitter.

B. Genmo Mochi 1 & Wan2.1

High-parameter open video models (10B+ parameters) that leverage the GB10's 90GB accessible VRAM ceiling, previously only runnable on enterprise cluster multi-GPU nodes.

---

Summary: The Single-Machine Autonomous Lab