Frontier open weights

The giants: open models you (probably) can't run at home.

Some of the most important open-weight models are simply too large for consumer hardware. They still matter: they set the benchmark ceiling, their licenses shape the ecosystem, and their distilled siblings are often the best models you can run. Here is what each one would actually take — and the realistic way to use it anyway.

Looking for something that runs on your machine? Try the hardware picker or browse the curated directory.

Kimi K3

Moonshot AI · China · Released July 16, 2026

Moonshot Open License (weights expected July 27, 2026)

The largest open-weight model ever released. Ranked fourth among all frontier models on independent testing, ahead of several closed flagships, and first in the Frontend Code Arena.

Size: 2.8T MoE (16 of 896 experts active, ~1.8%)
Context: 1M tokens, native multimodal
What running it yourself actually takes: Even at aggressive 4-bit quantization the weights alone exceed 1.4 TB. Running it means a multi-node GPU cluster — think 16+ datacenter GPUs with fast interconnect. This is not a homelab project; it is infrastructure.
Realistic access: Moonshot's API at $3 per million input tokens and $15 per million output. Cloud GPU rental for batch workloads. Several inference providers are expected to host it once weights land.
Runnable sibling: Kimi K2.7 Code (June 2026) — the agentic-coding sibling that quantized builds can run on a 96 GB+ workstation.

Official site →

GLM-5.2

Z.ai (Zhipu) · China · Released June 13, 2026

MIT

The current #1 open-weight model on the Artificial Analysis Intelligence Index. Beats GPT-5.5 on several long-horizon coding benchmarks at roughly one-sixth the price — under an MIT license.

Size: 744B
Context: 1M tokens
What running it yourself actually takes: Around 370–400 GB at 4-bit quantization. Technically within reach of a maxed-out Mac Studio cluster or a 4× H100 node, but far outside single consumer GPUs. Budget five figures for hardware that runs it acceptably.
Realistic access: Z.ai's API is aggressively priced. Most inference providers (Together, Fireworks, DeepInfra) host it. The MIT license means anyone can serve it, which keeps prices competitive.
Runnable sibling: GLM-4.7-Flash — the distilled fast variant that runs on a 24 GB GPU and stays in our trending list.

Official site →

Llama 4 Maverick

Meta AI · United States · Released 2025

Llama 4 Community License

Meta's frontier-tier open weight. The MoE design means inference speed comparable to a 17B dense model — if you can fit the full weights in memory.

Size: 400B MoE (~17B active)
Context: 1M tokens
What running it yourself actually takes: Roughly 200–230 GB at 4-bit. A 4× A6000 workstation or a 256 GB unified-memory Mac cluster gets you there; a single consumer GPU does not. The active-parameter trick helps speed, not memory.
Realistic access: Hosted by every major inference provider. Meta's own llama.com API. Often the cheapest frontier-class option per token because so many providers compete on it.
Runnable sibling: Llama 4 Scout — same family, 109B total, runs (tightly) on a 96 GB workstation and features in our picker.

Official site →

DeepSeek V4 Pro

DeepSeek AI · China · Released May 2026

MIT

The model that has dominated our trending list since launch: top open-weight scores on SWE-Bench Verified and GPQA Diamond, MIT licensed, and the default recommendation for serious code and math work.

Size: Large-scale MoE
Context: 1M tokens
What running it yourself actually takes: The full MoE needs several hundred gigabytes of memory across multiple GPUs. Community 4-bit builds exist but still demand workstation-cluster territory, not a desktop.
Realistic access: DeepSeek's own API is famously cheap. The MIT license means third-party hosting is plentiful and prices keep falling.
Runnable sibling: DeepSeek V4 Flash — the distilled variant. Quantized builds run on high-memory workstations, and it holds the #2 spot in our trending list.

Official site →