Downloads · 30 days
22
42% of all-time downloads
GLM-4.5-Air-midtrain-swe50-star50-90B
TianchenGuan/GLM-4.5-Air-midtrain-swe50-star50-90B
GLM-4.5-Air-midtrain-swe50-star50-90B is a machine learning model from TianchenGuan. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
Continued pretraining ("midtrain") of zai-org/GLM-4.5-Air-Base on a 50/50 mixture of SWE-agent-trajectory tokens (SWE-ZERO OSS trajectories + own distillation) and StarCoder-v2-anchored code / math / STEM / web text (…
All-time downloads
52
Public
Parameters
109B
218 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors218 GB · 100%
Parameter types
How the weights are stored.
BF16109B · 100%
Base models
Model card
From the Hugging Face model README
GLM-4.5-Air midtrain (swe50-star50, ~90B tokens)
Continued pretraining ("midtrain") of zai-org/GLM-4.5-Air-Base on a 50/50 mixture of
SWE-agent-trajectory tokens (SWE-ZERO OSS trajectories + own distillation) and
StarCoder-v2-anchored code / math / STEM / web text (dolmino mix, nemotron-cc-math).
184 train datasets; 16k sequence length; trained for 19,160 steps.
This is the initialization checkpoint used by the downstream SFT-v3 -> RL pipeline (the "bigrun"). At matched SFT steps, initializing from this checkpoint instead of the raw base improved SWE-bench-Verified by +6 to +9 points.
- Full data recipe (all 184 datasets and weights): see the accompanying configs release.
- Midtrain author: Wai Tong Chung. Released as part of a Duke University research project.
- Type
- glm4_moe
- License
- mit
- Created
- Aug 21, 2026
- Updated
- Aug 21, 2026