The open-weight landscape of August 2026

"Open-weight" is a label about the license, not about the hardware you need. The three most talked-about models this semester make that crystal clear.

Qwen3.6 — the one you can actually run yourself

Alibaba kept the most active release cadence of the semester: Qwen3.5 in February, Qwen3.6 in April. The 27B dense variant is the real sweet spot — it runs on one or two consumer GPUs in Q4 quantization, with 262K of native context and multimodal capability. It's, today, the most practical option for someone who wants real self-hosting without buying a rack.

Kimi K3 — open, but at datacenter scale

Moonshot AI released Kimi K3 in July: 2.8 trillion parameters in a MoE architecture, with 104B active per token and a context window of approximately 1M tokens. It is, by its own metrics, the largest open-source model released to date. But "released" doesn't mean "runnable on your desktop" — activating 104B parameters per token is still a load that demands datacenter infrastructure, not a consumer GPU.

DeepSeek V4 — the strongest efficiency story in the batch

DeepSeek reset its architecture toward hybrid attention in April: the V4-Pro variant reaches 1.6 trillion total parameters with 49B active, the V4-Flash variant stays at 284B total / 13B active. Even the "small" variant (Flash) requires on the order of 170GB of VRAM at Q4 quantization — out of reach of a single GPU, but feasible on a dedicated multi-GPU server.

The honest summary

If you have one or two GPUs and want to run something today: Qwen3.6 27B. If you want maximum open-weight capability and don't mind paying a provider to host it: Kimi K3 or DeepSeek V4 via API. The word "open" tells you what you can do with the license — not what you can run on your machine.