This week: HighVRAM expanded its large-memory local-AI planning coverage from a 96GB CUDA workstation through 192GB HBM and 512GB shared unified memory. The practical lesson is that capacity opens different model formats, but software support and measured workload behavior still decide whether a system is useful.
Delivery note: this is a web and RSS edition. HighVRAM email signup and delivery are not active, and publication here does not mean an email was sent.
What changed
The new Mac Studio M5 Ultra versus RTX PRO 6000 Blackwell comparison frames a common workstation decision: up to 512GB of shared unified memory with MLX/Metal support, or 96GB of dedicated GDDR7 ECC with a CUDA path. It does not treat memory capacity, bandwidth or AI TOPS as a universal speed ranking.
The dedicated AMD Instinct MI300X 192GB guide adds a third architecture: an OAM datacenter accelerator with 192GB HBM3, 5.3TB/s peak theoretical bandwidth and ROCm requirements. It is not a desktop PCIe card.
What the exact files say
| Qwen3.5 artifact | Complete files | First practical class in this week’s guides |
|---|---|---|
| 122B-A10B GPTQ-Int4 | 78.864GB | 96GB, with limited file-only headroom |
| 122B-A10B FP8 | 127.163GB | 192GB dedicated HBM or 256GB shared memory |
| 122B-A10B BF16 | 250.173GB | 512GB shared memory or supported multi-device placement |
| 397B-A17B GPTQ-Int4 | 235.708GB | 512GB shared memory or supported multi-device placement |
| 397B-A17B FP8 | 406.152GB | 512GB shared memory or supported multi-device placement |
| 397B-A17B BF16 | 806.796GB | 1TB planning class |
One useful action
If you operate large-model hardware, run one representative workload and submit a reproducible record through the expanded benchmark builder. It now captures multi-device topology, placement strategy, CPU/system-memory offload, batch/concurrency, input modality and separate peak device and system memory. A result with those fields is more useful than an unexplained token-rate screenshot.
Continue reading
- 96GB-to-512GB hardware-guide update
- Mac Studio M5 Ultra 512GB guide
- RTX PRO 6000 Blackwell 96GB guide
- Qwen3.5 96GB-to-1TB hardware paths
- Private browser watchlist and focused RSS feeds
Sources and limits
- Apple Mac Studio M5 technical specifications
- NVIDIA RTX PRO 6000 specifications
- AMD Instinct MI300X specifications
- Qwen3.5-122B official repository
- Qwen3.5-397B official repository
Published October 25, 2026. Hardware specifications are manufacturer-published, and model totals were summed from the cited official repositories. HighVRAM has not independently benchmarked every combination. File fit does not establish compatibility, speed, quality, reliability or value.