Open-Weight Models in Borg: GLM, Kimi, and GPT-OSS

Open-weight models make their weights available to download, while hosted model services provide access through a managed service. Borg lists open-weight models from GLM, Kimi, and gpt-oss-120b, but a Borg listing does not mean the weights run on your computer.

Why weights matter

Downloadable weights can give users more control over deployment than a managed endpoint. The gpt-oss-120b source says its Apache 2.0 license permits customization and local deployment. It describes text input and output, configurable reasoning effort, and agent-style tool capabilities.

Local use still requires compute and deployment work. Open weights do not automatically mean a model is small, free to operate, or easy to run. Model size, license, hardware needs, and task fit are separate questions.

“Open” is not a performance rating.

The source describes gpt-oss-120b as the larger family member and notes a smaller 20b sibling.

The model families

The GLM catalog includes several generations, from GLM-4.5 and GLM-4.5-Air to GLM-5, GLM-5.1, GLM-5.2, and GLM-5.3. Their descriptions focus on coding, tool use, and multi-stage agent work, with details that vary by model. GLM-5.3 has reasoning always on and adjustable effort according to its listing.

gpt-oss-120b is an open-weight text model from OpenAI. Its description highlights private self-hosted reasoning, fine-tuning, text-based tool workflows, and configurable reasoning experiments. The source entry does not publish a standard hosted API token rate.

Kimi's open-weight listings include Kimi K3 and Kimi K2.7 Code. K3 is described for engineering and knowledge-work agents, with long context and image and video understanding. K2.7 Code focuses on repository work, tool use, and sustained software implementation.

Benchmark examples and limits

Some models have vendor-reported scores. For gpt-oss-120b, listed results include 92.5 on AIME 2025 (high reasoning, no tools), 80.1 on GPQA Diamond (high reasoning, no tools), and 62.4 on SWE-bench Verified (high reasoning). OpenAI vendor evaluation paper.

GLM-5 reports 77.8 on SWE-bench Verified, 56.2 on Terminal Bench 2.0, 86.0 on GPQA-Diamond, and 92.7 on AIME 2026 I. GLM-5 vendor model card.

Kimi K3 reports 93.5 on GPQA Diamond, 67.5 on DeepSWE, and 77.8 on ProgramBench. Kimi K3 vendor evaluation.

Kimi K2.7 Code lists 62.0 on Kimi Code Bench v2 and 53.6 on Program Bench. Kimi K2.7 Code vendor model card.

Treat each score as evidence about its named evaluation, not a universal rank.

The tests may use different methods. Even a shared benchmark name does not prove versions, tools, or settings match. The model data has no benchmark results for every GLM or Kimi version.

Choosing a lane

Start with the task, not the open-weight label. For long software tasks, compare GLM and Kimi coding and agent descriptions. For text workflows where customization or local deployment matters, review gpt-oss-120b and its license. Kimi K3 is described as handling images and video, while gpt-oss-120b is listed for text input and output.

Borg lists each model as a lane, so you can select one directly or use Borg Fusion routing as described in the model entries. Check each listing for inputs, reasoning settings, license information, and reported benchmarks.

FAQ

What does open-weight mean?

It means the model weights are available to download, which can allow customization or local deployment depending on the model and its license.

Which open-weight models are listed in Borg?

The source data marks gpt-oss-120b, Kimi K3, and several GLM models as open-weight.

Are benchmark scores directly comparable?

No. Scores come from different evaluations and sources, so compare results only when the benchmark setup is the same.