Most “best open model” lists rank by benchmark. The more useful question for most people is “what can I run?” This guide sorts recent open-weight releases by hardware. Everything comes from posts I’ve already researched, with sources linked there. I haven’t run any of these. Most benchmark numbers are vendor-reported, and I mark the few independent ones.
Tier 1: a single 24 GB GPU or a well-equipped laptop
NVIDIA Nemotron 3.5 Bolt 74B-A9B. Uses native 2-bit compression to fit in 24 GB, runs through Ollama or LM Studio, and is on Hugging Face. NVIDIA claims first among open models in its class on PinchBench. Vendor claim only. See the Nemotron post.
JetBrains Mellum 2.1. 12B total, about 2.5B active, 131K context, Apache 2.0. You still need memory for all 12B parameters. JetBrains’ own numbers show it ahead on LiveCodeBench but behind a similar-size rival on SWE-bench Verified (47.0 vs 50.0). Built as a cheap engine for sub-agent work, not a top reasoner. See the Mellum post.
Tier 2: a workstation with a lot of memory, or one server GPU node
NVIDIA Nemotron 3.5 Super VL. 120B total, 12B active, OpenMDW license, 1M context. NVIDIA’s speed claim (about 625 output tokens/second) was measured on four B200 GPUs, so treat it as a server-class model. It’s aimed at agents that read documents and charts. The benchmark wins are on NVIDIA-chosen tasks with undisclosed details.
Tier 3: about 1 TB of memory, such as two top-spec Mac Studios
Meta Muse Spark Open. 816B total, 42B active, 1M context. Meta’s showcase is two Apple M5 Ultra Mac Studios, and TechCrunch cites at least 1 TB of RAM. The license is Meta’s custom Muse Model License, which per one analysis requires a separate agreement if you resell access above $20 million in revenue and requires “Built with Muse Spark Open” attribution, so read it before building a business on it. One independent number: Artificial Analysis scores it 39, ninth of 118 open models over 150B parameters. See the Muse Spark Open post.
Tier 4: data-center only
Mistral Large 4. About 1 trillion total parameters (52B or 49B active, sources differ), 1M context. It’s currently a preview through Mistral’s own API. Weights are promised by the end of October, and one estimate puts even a 4-bit version at about 500 GB of memory. The license is unpublished, and was reportedly not Apache 2.0. Don’t plan around it until the weights and license are out. See the Mistral post.
A quick chooser
| If you have… | Start with | Watch out for |
|---|---|---|
| One 24 GB GPU | Nemotron 3.5 Bolt 74B-A9B, Mellum 2.1 | Vendor-only benchmarks |
| A GPU server | Nemotron 3.5 Super VL | Task details undisclosed |
| ~1 TB unified memory | Muse Spark Open | Custom license terms |
| Data-center budget | Mistral Large 4 (when weights ship) | License and weights still pending |
| None of the above | Hosted APIs of the open models | Pricing and data policy |
If you don’t have the hardware, several of these are available through hosted APIs. Meta lists seven providers for Muse Spark Open, and Nemotron is on OpenRouter. That gives you the model without the memory bill, at the cost of sending your code to a provider.
How I’d test one
My suggestion, not a standard:
- Pick five real tasks from your own repo: a bug fix, a small feature, a test, a refactor and a doc update.
- Give each candidate the same prompt, the same agent harness and the same permissions.
- Judge by tests passing, not by how the answer reads.
- Record cost or time per accepted change.
- Repeat each task at least twice, because one run proves little.
For harness options, see OpenCode 2.0 and AgentPlane. If your agent runs untrusted code or reads untrusted pages, put it in a sandbox, as in the Perplexity SPACE post.
What I can’t tell you
Independent evaluations of these models are thin. For several of them the only published comparisons are the vendors’ own. Until outside tests catch up, a list like this tells you what fits on your machine and what the license allows, not which model is best.
Sources: the linked posts above, which cite NVIDIA coverage via AI Weekly, Meta, Hugging Face’s Mellum 2.1 model card and Let’s Data Science on Mistral Large 4.
