Last updated: October 17, 2026
For most small offices, a 24 GB GPU or a Mac with 32 GB or more of unified memory is the sweet spot in October 2026. That runs Qwen 3.8 27B and NVIDIA’s new Nemotron 3.5 Bolt locally through Ollama or LM Studio. A 16 GB machine handles small models only. Machines with 128 GB run 70B-class and 120B mixture-of-experts models.
Where things stand
The runtime question is mostly settled. Ollama, LM Studio and llama.cpp all run GGUF model files, and Ollama’s release notes update its bundled llama.cpp and MLX engines almost weekly. Ollama v0.40.0 (September 2026) made Apple’s MLX the default engine on Apple Silicon for supported models, so Macs got faster without any setup change. LM Studio stays the friendliest desktop app; its 0.4.17 release (June 2026) added support for AMD Strix Halo and Radeon AI PRO cards.
The 24 GB tier got a lot more interesting this week. Ollama v0.40.5 (October 14, 2026) added a “74B Nemotron 3.5 Bolt tuned for 24 GB GPUs.” Coverage of NVIDIA’s October 13 announcement describes it as a mixture-of-experts model with 9B active parameters, shipped at 2-bit with quantization-aware training. NVIDIA claims it beats similar-sized models, but no independent benchmarks were out when this page was written, and the license had not been confirmed.
Qwen 3.8 27B is the other default pick at this tier. Ollama’s build is about 18 GB, Apache 2.0 licensed, and needs a 24 GB GPU or a Mac with 24 GB or more. A guide from August 2026 is blunt that “16GB machines are below the waterline” for it. Long context windows add memory on top, so a 24 GB card can still run out if you raise the context length.
Google’s Gemma 4 (April 2026, Apache 2.0) covers the small end, with E2B, E4B, 12B, 26B mixture-of-experts and 31B dense sizes. On a 16 GB laptop, the small Gemma sizes are the honest choice. The large open models are not desk-side models. DeepSeek V4-Flash weights (MIT, July 2026) are about 167 GB, which puts them in multi-GPU territory. The full Qwen 3.8 release is a 2.4-trillion-parameter model that no consumer machine can hold.
Hardware prices are the real problem right now. The RTX 5090 lists at $1,999, but US street prices ran from about $4,300 in store to over $6,000 online in mid-September 2026 because of the memory shortage. NVIDIA’s DGX Spark (128 GB) rose from $3,999 at launch to $4,699 by October 2026. Unified-memory machines from Apple and NVIDIA’s new RTX Spark laptops are now the more predictable way to buy memory.
| Option | Best for | Cost (month seen) | Runs where | Our take |
|---|---|---|---|---|
| 16 GB laptop or Mac mini M6 (16 GB base) | Drafting, summaries, trying local AI | Mac mini M6 from $899 (Aug 2026) | One desk, Ollama or LM Studio | Small Gemma 4 sizes only. Fine for a pilot, not for a shared office assistant. |
| Mac mini M6, 32 GB | One user running a 27B model | Upgrade price not published in our sources | One desk | 32 GB is the Mac mini’s ceiling. Enough for Qwen 3.8 27B with modest context. |
| 24 GB GPU (RTX 4090 / 5090 class) | Qwen 3.8 27B, Nemotron 3.5 Bolt, a small team server | RTX 5090 $1,999 list, ~$4,300 to $6,400+ street (Sep 2026) | Windows or Linux tower, shared over the LAN | Fastest tier for the money if you can find a card. A used 4090 is the cheaper route. |
| Mac Studio M5 Max, 128 GB | 70B models, 120B MoE, quiet office server | From $2,499 (36 GB); about $5,399 reported for 128 GB (Sep 2026) | Desk or closet, macOS | 614 GB/s memory bandwidth, more than twice the DGX Spark’s. Bandwidth is what limits generation speed. |
| NVIDIA DGX Spark, 128 GB | CUDA tooling, 70B and 120B MoE models | $4,699 (Oct 2026) | Small Linux box on the network | Strong prompt processing, slower generation (273 GB/s). Good for developers, less so as a chat server. |
| RTX Spark laptops (N1X) | Portable 24 to 128 GB unified memory | No prices published (Sep 2026) | Laptop | Bolt was built for these. Wait for prices and reviews. |
What a small office should do now
If you want one shared local model for a 5 to 15 person office, buy memory first. A Mac Studio with 64 to 128 GB or a tower with a 24 GB GPU, running Ollama behind Open WebUI, covers drafting, summarizing and document Q&A on Qwen 3.8 27B. Pull Nemotron 3.5 Bolt next to it and compare on your own documents before you standardize.
Do not pay $5,000 or more for an RTX 5090 right now unless you already know what workload needs it. A used 4090 or a Mac with 64 GB or more gets most offices the same result.
Do not bother with local models at all if your work is not confidential and your team already pays for a cloud AI plan with business data terms. Local AI pays off when files cannot leave the building, not as a way to save on subscriptions.
What we’re watching
- Independent benchmarks and a confirmed license for Nemotron 3.5 Bolt, plus the Bolt E4B model slated for October 23.
- RTX Spark laptop prices and shipping dates from Lenovo, Dell, HP and others.
- Whether Ollama’s Qwen 3.8 lineup grows beyond the 27B size.
- RTX 5090 street prices as memory supply recovers.
- Mac Studio M5 Ultra 512 GB, due late October.
Sources
- Ollama release history (releases.sh), checked October 2026
- OrcaRouter: Nemotron 3.5 Bolt 74B release details, October 2026
- NVIDIA blog: local AI and Nemotron 3.5 Lightning, August 2026
- Yotta Labs: running Qwen 3.8 with Ollama, August 2026
- ModemGuides: Qwen 3.8 local hardware reality check, August 2026
- Wikipedia: Gemma (language model), checked October 2026
- Yotta Labs: DeepSeek V4 specs and access, August 2026
- LM Studio 0.4.17 changelog, June 2026
- Capital and Compute: RTX 5090 pricing, September 2026
- AI Indigo: DGX Spark 128 GB, October 2026
- AI Indigo: Mac Studio M5 Max 128 GB, October 2026
- AICybr: Mac Studio M5 and Mac mini M6 pricing, August 2026
- Implicator: RTX Spark October launch, no price, September 2026
Changelog
- 2026-10-17 — Page created.