Sources and verification dates.

Durable memory, power, and architecture specifications; board-partner dimensions vary.

Reference 32 GB, 575 W, and 1000 W system-power guidance.

Official 96 GB ECC, 1792 GB/s, 600 W workstation specification.

Official 32 GB, 640 GB/s, 300 W and Linux ECC specifications.

Current configurator availability may be narrower than durable platform maximums; catalog uses the currently observed 96 GB configuration.

Primary launch source for M3 Ultra architecture and unified-memory bandwidth.

CPU platform and product-family source; exact street prices remain estimates.

Primary platform source for high-lane-count multi-GPU workstation builds.

Compatibility reference for motherboard form factor and expansion slots.

Compatibility reference for multi-GPU workstation layouts.

PSU capacity and connector-family reference.

Chassis clearance and cooling-layout reference.

Primary model metadata: 30.5B total, 3.3B active, 48 layers, GQA, and native 256K context.

Primary model metadata: 32.5B parameters, 64 layers, 8 KV heads, and 131072-token context.

Primary metadata and provider-reported OpenHands SWE-bench Verified result; provider claim is not independently merged with other scores.

Primary model metadata for the 16B-total/2.4B-active MoE coding model.

Primary model card; general model retained as a quality-oriented non-coding comparison.

OpenAI-compatible server, GGUF quantization, CUDA/HIP/Metal and partial offload support.

Runtime compatibility and OpenAI-compatible serving guidance; deployments must validate exact model/quant combination.

Runtime support and platform-specific quantization limitations.

Apple Silicon inference and local OpenAI-compatible server reference.

Private overlay and SSH access guidance without router port forwarding.

Community-submitted primary benchmark outputs; useful only as comparable evidence with environment caveats.

Measured Qwen3 30B-A3B Q4 generation on RTX 4090 and RTX 5090; comparable, not exact, for the Coder checkpoint.

Community benchmark with raw JSONL and methodology for dense Qwen 32B across context and quantization.

Observed listing around 1380 USD; availability and seller pricing can change.

Retail observation; pricing varied by workstation and Max-Q edition.

Board-partner availability reference; catalog price is a wide observed range, not live pricing.

Observed configuration availability; checkout pricing and lead times can change.

Versioned observed on-demand rates; provider pricing and availability can change.

Used-price bands are deliberately wide; condition, warranty, cooler, and seller history materially affect value.