# GLM-5

GLM-5 er en modell fra Z.ai med åpne vekter, lansert 2026-02-12. Kontekstvinduet er 202 752 tokener. Maksimal utdatalengde er 131 072 tokener. Prisen er $1,00 per million inndatatokener og $3,20 per million utdatatokener.

- Utvikler: Z.ai
- Id: zai-org/glm-5
- Ifølge rapporter: 2026-06-21T18:57:49.000Z
- Lansert: 2026-02-12T00:00:00.000Z
- Observasjon: 2026-09-17T00:11:02.202Z
  - [dokumentasjonslenke](https://models.dev/#glm-5)
- Status: Lansert
- Vektversjon: 2026-02-11
  - [dokumentasjonslenke](https://huggingface.co/api/models/zai-org/GLM-5/paths-info/c183ef8c61faee82855eca1ed9bb3a9a7ce3b0b2)
- Kanonisk: https://alexgetman.com/nb/models/z-ai/glm-5-803/

## Spørsmål

### Hvor stort er kontekstvinduet til GLM-5?

Kontekstvinduet er 202 752 tokener. Maksimal utdatalengde er 131 072 tokener.

### Hva koster GLM-5?

Prisen er $1,00 per million inndatatokener og $3,20 per million utdatatokener. Bufrede inndata koster $0,20 per million tokener.

### Er vektene til GLM-5 offentlig tilgjengelige?

Ja. Z.ai publiserer vektene.

### Når ble GLM-5 lansert?

Den 2026-02-12.

## Tidslinje

- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · NVIDIA/TensorRT-LLM · @kaiyux](https://github.com/NVIDIA/TensorRT-LLM/pull/21019)
> - `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`.
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · NVIDIA/TensorRT-LLM · @moraxu](https://github.com/NVIDIA/TensorRT-LLM/pull/22136)
> 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`
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · ROCm/ATOM · @whx-sjtu](https://github.com/ROCm/ATOM/pull/3318)
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · ai-dynamo/dynamo · @indrajit96](https://github.com/ai-dynamo/dynamo/pull/18363)
> 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
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · ai-dynamo/frontend-crates · @dundysm](https://github.com/ai-dynamo/frontend-crates/pull/600)
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · anomalyco/models.dev · @models-dev-automation[bot]](https://github.com/anomalyco/models.dev/pull/13561)
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · maximhq/bifrost · @mickgvirtu](https://github.com/maximhq/bifrost/pull/4593)
> 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
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · milabench/milabench · @bouthilx](https://github.com/milabench/milabench/pull/65)
> - GLM-5: layout check before launch, loss from the last stage, checkpoint timing and checksums, crash cleanup; `benchmarks/megatron/dev.yaml`, `scripts/trillium/glm/`.
- 2026-11-13T19:01:50.882Z · [GitHub search · Z.ai · pulls · smg-project/smg · @hello-alexmcc](https://github.com/smg-project/smg/pull/6229)
> 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

## Dokumentasjonslenker

- [vertex-model-garden](https://console.cloud.google.com/vertex-ai/publishers/zai-org/model-garden/glm-5)
- [review](https://docs.z.ai/api-reference/llm/chat-completion)
- [zai-pricing](https://docs.z.ai/guides/overview/pricing)
- [GitHub search · Z.ai · pulls · NVIDIA/TensorRT-LLM · @kaiyux](https://github.com/NVIDIA/TensorRT-LLM/pull/21019)
- [GitHub search · Z.ai · pulls · NVIDIA/TensorRT-LLM · @moraxu](https://github.com/NVIDIA/TensorRT-LLM/pull/22136)
- [GitHub search · Z.ai · pulls · ROCm/ATOM · @whx-sjtu](https://github.com/ROCm/ATOM/pull/3318)
- [GitHub search · Z.ai · pulls · ai-dynamo/dynamo · @indrajit96](https://github.com/ai-dynamo/dynamo/pull/18363)
- [GitHub search · Z.ai · pulls · ai-dynamo/frontend-crates · @dundysm](https://github.com/ai-dynamo/frontend-crates/pull/600)
- [GitHub search · Z.ai · pulls · anomalyco/models.dev · @models-dev-automation[bot]](https://github.com/anomalyco/models.dev/pull/13561)
- [GitHub search · Z.ai · pulls · maximhq/bifrost · @mickgvirtu](https://github.com/maximhq/bifrost/pull/4593)
- [GitHub search · Z.ai · pulls · milabench/milabench · @bouthilx](https://github.com/milabench/milabench/pull/65)
- [GitHub search · Z.ai · pulls · smg-project/smg · @hello-alexmcc](https://github.com/smg-project/smg/pull/6229)
- [hf-author-artifacts:zai-org](https://huggingface.co/api/models/zai-org/GLM-5/paths-info/c183ef8c61faee82855eca1ed9bb3a9a7ce3b0b2)
- [huggingface-router](https://huggingface.co/zai-org/GLM-5)
- [models-dev](https://models.dev/#glm-5)

## Signaler