This model was published in HF papers on 2026-02-17 and contributed to Hugging Face Transformers on 2026-02-09.

</div>

GlmMoeDsa

GlmMoeDsa (GLM-5) is a 744B-parameter mixture-of-experts model with 40B active parameters per token, using DeepSeek Sparse Attention (DSA) for efficient 200K-token context handling. It was trained entirely on Huawei Ascend chips and matches frontier-level performance on reasoning and long-context benchmarks.

The example below demonstrates how to generate text with [Pipeline] or the [AutoModelForCausalLM] class.

from transformers import pipeline


pipe = pipeline(
    task="text-generation",
    model="zai-org/GLM-5",
)
pipe("The theory of relativity states that")

from transformers import AutoModelForCausalLM, AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("zai-org/GLM-5")
model = AutoModelForCausalLM.from_pretrained(
    "zai-org/GLM-5",
    device_map="auto",
)
input_ids = tokenizer("The theory of relativity states that", return_tensors="pt").to(model.device)

output = model.generate(**input_ids, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))

[!NOTE] The MLA query LoRA path (q_lora_rank) is required. Like DeepSeek-V3.2, the DSA indexer scores queries from the low-rank query latent q_a_layernorm(q_a_proj(x)) (its wq_b projection is sized by q_lora_rank), so the model always uses the LoRA query path and q_lora_rank must be set — the released checkpoint uses 2048. The optional non-LoRA q_proj path that DeepSeek-V3 exposes for q_lora_rank=None is not supported here: without the query latent there is nothing for the indexer to consume.

GlmMoeDsaConfig

[[autodoc]] GlmMoeDsaConfig

GlmMoeDsaModel

[[autodoc]] GlmMoeDsaModel - forward

GlmMoeDsaForCausalLM

[[autodoc]] GlmMoeDsaForCausalLM - forward