Kimi K3, GLM-5.3 top Fastino's open-weight rankings
A Fastino guide calls Moonshot's Kimi K3 and Z.ai's GLM-5.3 the strongest downloadable models of summer 2026, now within three index points of Claude Opus 5.
Fastino's summer 2026 guide to open weight models, written by Thao Nguyen with data as of September 1, rates Kimi K3 and GLM-5.3 as a tie for agentic coding and reasoning. According to the post, both score 60 on the Artificial Analysis Intelligence Index, three points behind Claude Opus 5.
The two trade wins on individual benchmarks. GLM-5.3 leads on GDPval-AA v2 (1758 to 1668). Kimi K3 leads on Terminal-Bench v2.1 (85.0 to 83.9) and GPQA Diamond (93.5 to 91.7). The bigger difference is cost. Kimi K3 is a 2.8T-parameter model with 104B active parameters, priced at $3.00/$15.00 per million tokens and generating about 38 tokens per second. Its license requires a separate commercial agreement once a company's annual revenue passes $20 million. GLM-5.3 has 753B parameters (40B active) and costs $1.40/$4.40 on Z.ai's API. Fastino describes its license as near-MIT.
Fastino calls GLM-5.3-Flash, a 320B MoE model with 18B active parameters, the summer's most notable release. It scores 63% on DeepSWE against GLM-5.3's 69% at roughly one-sixteenth the cost. It first appeared anonymously as "Ox Alpha" and, per the post, became OpenRouter's biggest single-model launch before Z.ai revealed it on August 26.
The guide's other picks:
- Qwen3.8-27B (Apache 2.0), for fine-tuning and local use. It scores 52 on the index and runs on a single RTX 4090 at 4-bit. Citing Hugging Face data, Fastino says Qwen has about 151K derivative models.
- Thinking Machines Lab's Inkling, as the best American option. It is a 975B-parameter Apache 2.0 model that scores 42.
The figures are Fastino's compilation and were not independently verified. The available text ends partway through the Inkling section.