SparkLab 0.1.2

SparkLab 0.1.2

SparkLab 0.1.2 is a speculative-decoding performance release for NVIDIA DGX Spark. It packages the optimized native MTP, DFlash2, and DSpark paths merged since 0.1.1.

Highlights

• Qwen3.6-35B-A3B MTP2 reaches 74.42 tok/s. The optimized width-two path is 64.0% faster than its matched 45.38 tok/s eager control. MTP remains opt-in; Fast certification continues to apply to the target-only profile.
• Qwen3.8-Flash-Next MTP3 reaches 30.67 tok/s with 0.258 s warm TTFT. That is a 55.35% improvement over the controlled target run, with the same 128-token output hash across all selected trials.
• GLM-5.3 Flash MTP3 reaches 7.44 tok/s. The optimized path uses asynchronous expert prefetch, averages 87.9% draft acceptance, and preserves exact target-only output parity.
• Qwen3.8-27B DFlash2-8 reaches 35.48 tok/s with 0.153 s warm TTFT. The Fast-tier experimental profile is 3.87x faster than its matched target-only control, preserves exact greedy output parity, and eliminates rejection replay.
• DeepSeek V4 DSpark reaches 13.15 tok/s. Small-row kernels, residency tuning, and confidence-aware speculation improve the current experimental profile.

Runtime improvements

• Native MTP support and optimized short-batch verification for current Qwen and GLM recipes.
• Native DFlash2 support for dense Qwen3.8-27B, including radix top-k selection, a fused selector walk, proposal-only NVFP4 projection, and transactional GDN state commits.
• Replay-free speculative verification and tighter KV/state-pool lifecycle handling.
• Versioned GB10 evidence for every performance claim above.

All accelerated profiles remain opt-in where documented. Experimental results do not promote a recipe to Certified without its complete context, capability, concurrency, and endurance suite.

Installation and upgrade

pip install --upgrade "sparklab[accel]==0.1.2"

Install the attached sparklab_kernel_cache-0.1.2+cu130-py3-none-linux_aarch64.whl alongside the runtime. Runtime and kernel-cache release versions must match.

Requirements

• NVIDIA DGX Spark / GB10 with 128 GB coherent unified memory
• ARM64 Linux or DGX OS
• NVIDIA driver r580 or newer
• CUDA 13 toolkit
• Local NVMe storage for model and FTW artifacts

Validation and provenance

• Local suite: 1,692 passed, 8 skipped
• Pull-request CPU, DCO, and CodeQL checks passed
• TestPyPI rehearsal passed for CPython 3.10-3.13
• Production PyPI wheel matrix, GitHub build attestations, and SHA-256 checksums passed
• Signed immutable tag: v0.1.2

Documentation

• Installation
• Model catalog
• GB10 benchmark evidence
• PyPI package

SparkLab is developed by SixteenMiles Labs, a research lab under Oakmind AI.

SparkLab 0.1.1

SparkLab 0.1.1

SparkLab 0.1.1 is a performance and evidence release for NVIDIA DGX Spark. It packages the reviewed optimization, certification, compatibility, and benchmark work merged since 0.1.0.

Highlights

• Qwen3.6-35B-A3B is now Fast-tier Certified. The pinned NVFP4 FTW recipe measured 67.79 decode tok/s and 0.329 s warm TTFT, passed exact 32K recall and capability probes, and completed a zero-swap 60-minute endurance run.
• Qwen3.8-Flash-Next reaches 16.61 decode tok/s. Dense batch-one QSA now replays a CUDA graph and automatically falls back to eager sparse QSA beyond the dense budget. Exact 64K recall and reasoning, tool, coding-agent, and quality probes passed.
• GLM-5.3 Flash reaches 6.27 decode tok/s. Fused mHC improves decode throughput by 25.6% over the matched eager control.
• GLM-5.3 NVFP4 research path and FTW artifact. The release adds a bounded GB10 recipe and immutable artifact metadata for the full model.
• Safer runtime/kernel pairing. sparklab doctor now detects incompatible SparkLab and kernel-cache release or build stamps before model execution.
• Reproducible benchmark suites. Cross-framework Qwen3.6 results and W1-W4 paper-quality harnesses are included with machine-readable evidence.

Model status

Recipe	Status	GB10 result
Qwen3.6-35B-A3B NVFP4 FTW	Certified / Fast	67.79 tok/s, 0.329 s warm TTFT
Qwen3.8-Flash-Next NVFP4 FTW	Experimental / Frontier	16.61 tok/s, 0.403 s warm TTFT
GLM-5.3 Flash NVFP4/KDA-FP8 FTW	Experimental / Frontier	6.27 tok/s, 5.681 s warm TTFT
GLM-5.3 NVFP4 FTW	Experimental / Research	0.81 tok/s, 2.530 s warm TTFT

Model status remains evidence-bound. Qwen3.8-Flash-Next and both GLM-5.3 recipes have not completed every certification gate. Qwen3.6 certification establishes serving correctness and operational behavior; its fixed five-problem AIME sample exhausted the 2,048-token reasoning cap before producing final answers.

Installation and upgrade

pip install --upgrade "sparklab[accel]==0.1.1"

Install the attached sparklab_kernel_cache-0.1.1+cu130-py3-none-linux_aarch64.whl alongside the runtime. SparkLab 0.1.1 rejects a mismatched 0.1.0 kernel-cache package. Existing prepared FTW artifacts remain compatible when their recipe fingerprint and pinned revision match; no API migration is required.

Requirements

• NVIDIA DGX Spark / GB10 with 128 GB coherent unified memory
• ARM64 Linux or DGX OS
• NVIDIA driver r580 or newer
• CUDA 13 toolkit
• Local NVMe storage for model and FTW artifacts

Validation and provenance

• Hosted CPU suite: 814 passed, 20 skipped
• TestPyPI rehearsal: CPython 3.10-3.13 wheel matrix and strict twine checks passed
• Clean GB10 rehearsal-wheel smoke: Qwen3.6 measured 68.20 decode tok/s and 0.325 s warm TTFT
• Protected release workflow: trusted PyPI publishing, GitHub build attestations, and SHA-256 checksums

Documentation

• Installation
• Model catalog
• CLI reference
• GB10 benchmark evidence
• Cross-framework benchmark
• Quality benchmark
• PyPI package

SparkLab is developed by SixteenMiles Labs, a research lab under Oakmind AI.

SparkLab rolling beta

Automated rolling wheels from SixteenMiles Labs. Assets may be replaced in place.

SparkLab 0.1.0 Beta 2

SparkLab 0.1.0 Beta 2

SparkLab runs frontier open-weight models privately on one NVIDIA DGX Spark. Beta 2 is the first release built from the public SixteenMiles Labs repository, with public build provenance and hardened API error handling.

Highlights

• Native NVIDIA GB10 (SM121) support for ARM64 Linux and DGX OS
• CPython 3.10–3.13 manylinux ARM64 runtime wheels
• CUDA 13 kernel-cache wheel for reproducible GB10 startup
• Prebuilt FTW acquisition with pinned revisions and artifact validation
• RAM-resident and bounded NVMe-backed MoE execution
• OpenAI Chat Completions and Responses APIs, plus Anthropic Messages
• Sanitized daemon and API errors that preserve stable client classifications without exposing backend exception details
• Public CodeQL coverage for Python, C/C++, and GitHub Actions, with zero open findings at release time
• Public GitHub build attestations and PyPI trusted-publisher provenance

Model portfolio

• Qwen3.6-35B-A3B NVFP4 FTW — Fast-tier experimental recipe; measured at 67.46 decode tok/s and 0.320 s warm TTFT
• Qwen3.8-Flash-Next NVFP4 FTW — Frontier-tier experimental recipe; measured at 12.58 decode tok/s and 0.786 s warm TTFT
• DeepSeek V4 Flash DS-FP4 — Frontier-tier preview recipe; measured at 10.28 decode tok/s and 0.604 s warm TTFT
• GLM-5.3 Flash NVFP4/KDA-FP8 FTW — Frontier-tier experimental recipe; measured at 4.98 decode tok/s and 5.379 s warm TTFT
• Kimi K3 NVFP4/FP8 FTW — Research-tier experimental recipe with bounded NVMe execution

Model status is evidence-bound. Preview and Experimental recipes have not passed every context, correctness, parser, agent, and endurance gate required for Certified status.

Installation

pip install "sparklab[accel]==0.1.0b2"

The attached sparklab_kernel_cache-0.1.0b2+cu130-py3-none-linux_aarch64.whl provides the CUDA 13 kernel cache. Verify downloaded assets with SHA256SUMS.

Requirements

• NVIDIA DGX Spark / GB10 with 128 GB coherent unified memory
• ARM64 Linux or DGX OS
• NVIDIA driver r580 or newer
• CUDA 13 toolkit
• Local NVMe storage for models and FTW artifacts

Validation and provenance

The protected public release workflow built and checked the CPython 3.10–3.13 ARM64 matrix and published runtime wheels to PyPI through trusted publishing. GitHub and PyPI provenance bind the artifacts to this repository, workflow, tag, and release commit.

Documentation

• Installation
• Model catalog
• CLI reference
• GB10 benchmark evidence
• PyPI package

SparkLab is developed by SixteenMiles Labs, a research lab under Oakmind AI.

SparkLab 0.1.0

SparkLab 0.1.0

SparkLab runs frontier open-weight models privately on one NVIDIA DGX Spark. Version 0.1.0 is the first stable Python package release from the public SixteenMiles Labs repository.

Highlights

• Native NVIDIA GB10 (SM121) support for ARM64 Linux and DGX OS
• CPython 3.10–3.13 manylinux ARM64 runtime wheels
• CUDA 13 kernel-cache wheel for reproducible GB10 startup
• Prebuilt FTW acquisition with pinned revisions and artifact validation
• RAM-resident and bounded NVMe-backed MoE execution
• OpenAI Chat Completions and Responses APIs, plus Anthropic Messages
• Sanitized daemon and API errors with stable client classifications
• Public CodeQL coverage with zero open findings at release time
• Public GitHub build attestations and PyPI trusted-publisher provenance

Model portfolio

• Qwen3.6-35B-A3B NVFP4 FTW — Fast-tier experimental recipe; measured at 67.46 decode tok/s and 0.320 s warm TTFT
• Qwen3.8-Flash-Next NVFP4 FTW — Frontier-tier experimental recipe; measured at 12.58 decode tok/s and 0.786 s warm TTFT
• DeepSeek V4 Flash DS-FP4 — Frontier-tier preview recipe; measured at 10.28 decode tok/s and 0.604 s warm TTFT
• GLM-5.3 Flash NVFP4/KDA-FP8 FTW — Frontier-tier experimental recipe; measured at 4.98 decode tok/s and 5.379 s warm TTFT
• Kimi K3 NVFP4/FP8 FTW — Research-tier experimental recipe with bounded NVMe execution

Model status remains evidence-bound. Preview and Experimental recipes have not passed every context, correctness, parser, agent, and endurance gate required for Certified status.

Installation

pip install "sparklab[accel]==0.1.0"

The attached sparklab_kernel_cache-0.1.0+cu130-py3-none-linux_aarch64.whl provides the CUDA 13 kernel cache. Verify downloaded assets with SHA256SUMS.

Requirements

• NVIDIA DGX Spark / GB10 with 128 GB coherent unified memory
• ARM64 Linux or DGX OS
• NVIDIA driver r580 or newer
• CUDA 13 toolkit
• Local NVMe storage for models and FTW artifacts

Validation and provenance

The protected public release workflow built and checked the CPython 3.10–3.13 ARM64 matrix and published runtime wheels to PyPI through trusted publishing. GitHub and PyPI provenance bind the artifacts to this repository, workflow, tag, and release commit.

Documentation

• Installation
• Model catalog
• CLI reference
• GB10 benchmark evidence
• PyPI package

SparkLab is developed by SixteenMiles Labs, a research lab under Oakmind AI.

v0.1.0b1

SparkLab 0.1.0 Beta 1

SparkLab runs frontier open-weight models privately on one NVIDIA DGX Spark. This first beta packages the GB10-native runtime, immutable model recipes, FTW checkpoint support, unified-memory admission, and local OpenAI- and Anthropic-compatible APIs.

Highlights

• Native NVIDIA GB10 (SM121) support for ARM64 Linux and DGX OS
• CPython 3.10–3.13 manylinux ARM64 runtime wheels
• CUDA 13 kernel-cache wheel for reproducible GB10 startup
• Prebuilt FTW acquisition with pinned revisions and artifact validation
• RAM-resident and bounded NVMe-backed MoE execution
• OpenAI Chat Completions and Responses APIs, plus Anthropic Messages
• Memory, swap, dependency, storage, and platform checks through sparklab doctor
• Versioned GB10 benchmark evidence and model-specific run instructions

Model portfolio

• Qwen3.6-35B-A3B NVFP4 FTW — Fast-tier experimental recipe; measured at 67.46 decode tok/s and 0.320 s warm TTFT
• Qwen3.8-Flash-Next NVFP4 FTW — Frontier-tier experimental recipe; measured at 12.58 decode tok/s and 0.786 s warm TTFT
• DeepSeek V4 Flash DS-FP4 — Frontier-tier preview recipe; measured at 10.28 decode tok/s and 0.604 s warm TTFT
• GLM-5.3 Flash NVFP4/KDA-FP8 FTW — Frontier-tier experimental recipe; measured at 4.98 decode tok/s and 5.379 s warm TTFT
• Kimi K3 NVFP4/FP8 FTW — Research-tier experimental recipe with bounded NVMe execution

Model status is evidence-bound. Preview and Experimental recipes have not passed every context, correctness, parser, agent, and endurance gate required for Certified status.

Installation

Install the runtime from PyPI:

pip install "sparklab[accel]==0.1.0b1"

The attached sparklab_kernel_cache-0.1.0b1+cu130-py3-none-linux_aarch64.whl provides the CUDA 13 kernel cache. Verify downloaded assets with SHA256SUMS.

Requirements

• NVIDIA DGX Spark / GB10 with 128 GB coherent unified memory
• ARM64 Linux or DGX OS
• NVIDIA driver r580 or newer
• CUDA 13 toolkit
• Local NVMe storage for models and FTW artifacts

Validation

The release workflow built and checked the CPython 3.10–3.13 ARM64 matrix, published runtime wheels to PyPI through trusted publishing, and produced the attached checksums. The exact PyPI wheel was also exercised against a complete Qwen3.6 NVFP4 FTW checkpoint through the streaming API with zero scoped swap growth and zero OOM markers.

Documentation

• Installation
• Model catalog
• CLI reference
• GB10 benchmark evidence
• PyPI package

SparkLab is developed by SixteenMiles Labs, a research lab under Oakmind AI.