GLM-5

时间线

1. 2026-02-11 · 权重版本 · 参考链接

1. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · smg-project/smg · @hello-alexmcc
   摘录

   Atria-Dawn, `glm`, beside Atria-Dawn-Preview: the template writes the GLM-5 shape (`[gMASK]<sop>`, a reasoning-effort system turn, `<|system|>` / `<|user|>` / `<|assistant|>` / `<|observation|>` turns), the tools

2. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · milabench/milabench · @bouthilx
   摘录

   - GLM-5: layout check before launch, loss from the last stage, checkpoint timing and checksums, crash cleanup; `benchmarks/megatron/dev.yaml`, `scripts/trillium/glm/`.

3. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · maximhq/bifrost · @mickgvirtu
   摘录

   is not `schemas.Anthropic`, and the served model -- `glm-5`, `Kimi-K2`, ... -- is neither opus-4.8 nor Fable), so Claude Code's per-turn reminders were hoisted every

4. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · anomalyco/models.dev · @models-dev-automation[bot]
5. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · ai-dynamo/frontend-crates · @dundysm
6. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · ai-dynamo/dynamo · @indrajit96
   摘录

   for any DSA model (including DeepSeek V3.2 and GLM-5), or matching `flashmla_kv`/`trtllm` kernels on K3.1's sharded DCP path. Kernels SGLang has not resolved at CLI

7. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · ROCm/ATOM · @whx-sjtu
8. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · NVIDIA/TensorRT-LLM · @moraxu
   摘录

   class is registered for both `DeepseekV3ForCausalLM` and `DeepseekV32ForCausalLM`; GLM-5 (`GlmMoeDsaForCausalLM`) and Mistral Large 3 subclass it, so exact-class matching excludes them. Kimi K2 declares `DeepseekV3ForCausalLM`

9. 2026-11-13T19:01:50.882Z · 参考链接 · GitHub search · Z.ai · pulls · NVIDIA/TensorRT-LLM · @kaiyux
   摘录

   - `modeling_glm5_next_vision.py`: the GLM-5 input processor takes the prefix tokenization cache (the fourth commit). Before, it tokenized every text prompt in full and ignored `enable_tokenization_cache`.