GLM-5.3 Flash auf OpenRouter gesichtet
Durch Belege gestütztOffene Modellgewichte25 Beleglinks
Englisches Original

Z.ai
USD / 1 Mio. Token
| Eingabe | Ausgabe | Cache-Eingabe |
|---|---|---|
|
|
|
| Modell | Beleglinks | Preis Ein-/Ausgabe |
|---|---|---|
glm-5.3-flashstandard · Text | opencode-go | $0,15 / $0,50 |
glm-5.3-flashstandard | zai-pricing API-Route nicht verfügbar | $0,15 / $0,50 |
Agent-Model: glm-5.3-flash Agent-Effort: high
the B5/v4 implementation and all integrations by actual ollama-cloud/glm-5.3-flash sessions (opencode-db safe metadata in package1-b5-v2-20261028/provenance/finish-phase-model-metadata.json); the earlier reviews were actually GPT (openai/gpt-5.6-sol high: windows-strict 520c2edd…,
z-ai/glm-5.3-flash as fallback, and a 404 'no provider serves this model'
build (deepseek-v4.1-flash) -> audit (qwen3.8-flash) -> review (glm-5.3-flash as mandated) -> design (gpt-6-luna). The auditor is a different model family
Model: opencode:zai-coding-plan/glm-5.3-flash@high
Model: opencode:zai-coding-plan/glm-5.3-flash@high
Model: opencode:zai-coding-plan/glm-5.3-flash@high
breaker then cooled every healthy sibling model — glm-5.3-flash never failed yet was blocked by is_cooling_for ORing
After the global LLM connector moved to llmbase.ai (z-ai/glm-5.3-flash), every tool call
Generated-By: pi 1.5.2 (glm-5.3-flash @ zhipuai-coding-plan, reasoning=max)
Generated-By: pi 1.5.2 (glm-5.3-flash @ zhipuai-coding-plan, reasoning=max)
Generated-By: pi 1.5.2 (glm-5.3-flash @ zhipuai-coding-plan, reasoning=max)
Generated-By: pi 1.5.2 (glm-5.3-flash @ zhipuai-coding-plan, reasoning=max)
json.providers.nebius.models["zai-org/GLM-5.3-Flash"].reasoning.effort[2] = "max";
GLM-5.3: the text model of GLM-5.3-Flash (glm5_next) and GLM-5.3 (glm_moe_dsa),
auto: iteration 1-bug [model: nvidia/z-ai/glm-5.3-flash]
by ModelOpt itself (NVIDIA/Model-Optimizer#3176), large MoE models (e.g. GLM-5.3-Flash), lm-eval accuracy, layerwise weight reload, EPLB, experts with biases.
(open) edits the deprecated shared sizes. #63070 (draft, GLM-5.3-Flash) and #66320 (Qwen3.5) would move more models to breakable graphs on gfx950, which then get these
compressed scale format (CSF) in GPU memory. On GLM-5.3-Flash with TP2, each rank uses 3.53 GiB less memory, and the KV cache holds 30.6% more
In conjunction with #60229, support adaptive verification for GLM-5.3-Flash on Hopper GPUs by overriding `FlashInferMLASparseSM90Builder.get_varlen_cudagraph_max_query_len`.
[LoRA][Model] Add LoRA support for GLM-5.3-Flash <!-- markdownlint-disable -->
no longer fit next to the running requests. GLM-5.3-Flash NVFP4 on 2x B200 (TP2, the 0.9 split that was its default before #50018: 2,969,600 KV
mHC models (the DeepSeek-V4 family, GLM-5.3-Flash) run the unfused mHC path: a chain of small kernels
On GLM-5.3-Flash with attention DP8 (one TPU v7x 2x2x1 host, AgentX trace replay), round robin caches 0.22 of prompt tokens at c16, against a theoretical
- **#84** is for `z-ai/glm-5.3-flash`: the same key and the same launcher our verifiers log, with release marks.
200 on 2026-11-13 (`nemotron-3-ultra`, `nemotron-3.5-super-vl`, `nemotron-3.5-lightning`, `nemotron-3.5-bolt`, `minimax-m3.1-flash`, `glm-5.3-flash`). Default is `nvidia/nemotron-3-ultra-550b-a55b`. Retired ids are recorded in the file as a comment with EOL dates,
| router softmax, 288 experts (GLM-5.3-Flash) | 16 tokens | 1.75x | 1.01x | 1.7x at 16384 tokens (14.1 to 8.1 us) |
separate reviewer session on the same model family (GLM-5.3-Flash, no model diversity available). Four findings fixed: ownership cross-validation lives in `charts/courier/templates/common.yaml.tpl` (not `secure.yaml.tpl`), isolation test
| `z-ai/glm-5.3-flash` | reasoning, fallback | NEAR AI-hosted reasoning model; too slow for the 30s chat-stream timeout |
showed prices. `cn:deepseek-v4.1-flash` came from OpenRouter (0.3/1.2). `cn:kimi-k3.1-pro`, `cn:glm-5.3-flash`, `zd/gpt-6-sol` and `zd/gemini-4-argon` came from models.dev.
Created on behalf of Ifiok Jr. (@ifiokjr), GLM-5.3-Flash (zai-coding-plan/glm-5.3-flash), high reasoning.
claude-code 71, codex-cli 46, antigravity 6, openai-oauth 67, glm-5.3-flash 12, glm-5.3 12.
logs de produção mostram interrupções por `TimeoutError` no `z-ai/glm-5.3-flash`; houve também HTTP 502 no endpoint separado `qa-ai`.
(`harbor run`) with `gemini/gemini-3.1-pro-preview` via `gemini-cli` (5 trials), `zai-org/GLM-5.3-Flash` via `terminus-2` (5 trials), and `Qwen/Qwen3.8-27B` via `terminus-2` (5 trials), plus Oracle (`1.0` reward, 4/4 pytest
necessary step for running large MoE models (e.g. GLM-5.3-Flash 320B, 157 GB GGUF) on consumer hardware (32 GB RAM, single GPU) where CPU offloading makes
Vision tower alone (glm-5.3-bf16 weights, identical to GLM-5.3-Flash's BF16 tower), 1x B200, eager as served by default, FA4. Two interleaved runs per variant, mean of
Anthropic stays at 20,000 from the provider layer; `z-ai/glm-5.3-flash` gets 131,072 from the model layer.
serve packs an outside quant plugin claims (DeepSeek-V4.1, GLM-5.3-Flash, Qwen3.8, Ling-3.1-flash)
Model: z-ai/glm-5.3-flash
On four GB10s at TP=4 (GLM-5.3-Flash XSFP4, `--context 1048576 --parallel 8`), round time in ms from our own grid client (`"draft": false`, greedy, 256 tokens,
as they are. On four GB10s at TP=4 (GLM-5.3-Flash XSFP4, `--context 1048576 --parallel 8`), round time in ms from our own grid client (`"draft": false`,
v1 (2; new models should usually be v2_only): isoquant/glm-5.3-flash, isoquant/deepseek-v4.1-flash-1017-fast
`z-ai/glm-5.3-flash`, `stepfun-ai/step-5-preview`, `meta/muse-glimmer-30b`, `nvidia/nemotron-3.5-super-vl-120b-a12b` e `nvidia/nemotron-3.5-lightning-30b-a3b`. O app consulta o catálogo real da conta e só tenta os IDs retornados quando há correspondência.
json.providers.nebius.models["zai-org/GLM-5.3-Flash"].reasoning.effort[2] = "max";
through the real TUI in a PTY with `opencode-go/glm-5.3-flash`. I captured the provider request bodies and read the SQLite rows instead of trusting the model's
- With the real model (glm-5.3-flash) in the TUI, one temporary home and two projects:
- With the real model (glm-5.3-flash) in the TUI, using temporary homes: a global Program ran by name from another project and recorded its exact
- Real model over a PTY (`opencode-go/glm-5.3-flash`) in a temporary home, since deleted. In a first session I stated a codename and a rollback window.
- Callers that reach the new default: GLM-5.3-Flash at attention TP 2 (32 heads per rank), bf16 KV cache, gfx950,
every shape. This is what SGLang passes for GLM-5.3-Flash under CUDA graphs: the k-pool indexer's block-table capacity, 1M context / kpool 4 = 4,097 pages
Draft for #6186: MODEL-AB-REGRESSION-1: qnfo-ai glm-5.3-flash -> standard on grounded-qa: win rate 0 < 0.4
(166 providers); `GET /api/llm/models?provider=wally` returns 10 models, including `wally/glm-5.3-flash` and `wally/deepseek-v4.1-flash`.
The [GLM-5.3-Flash model card](https://huggingface.co/zai-org/GLM-5.3-Flash) documents `reasoning_effort` and `clear_thinking` and no way to turn thinking off. No revision of these three repositories' templates has contained `enable_thinking`.
Add pooled DSA forward APIs for Blackwell GLM-5.3-Flash pools index keys before selecting attention tokens. Add forward APIs for gated cache updates, paged scoring, pool
end in #22082: a one-token prompt on the GLM-5.3-Flash FP8 target must not depend on the KDA slot's previous request, and the target is not
- **GLM-5.3-Flash** (`models/glm5_next/modeling_ptq_vllm.py`, a new package): the k-pool indexer with its K-cache writers and DeepGEMM scorer, from `vllm_indexer.py`. Like `kimi_k3`, `glm5_next` has no `specs.py`: it
fix(vite): release notes default to Gemma 4 before glm-5.3-flash retires (4.5.1)
- Changes the default model to `@cf/zai-org/glm-5.3-flash`, which followed the prompt rules more reliably than the previous default in testing. `RELEASE_NOTES_MODEL` still overrides it.
Move the Fortanix GLM 5.3 Flash offer from the homepage to its model page and search text
Model: zai-org/GLM-5.3-Flash Harness: Hermes Agent
submitted game. The game model remains the challenge's `z-ai/glm-5.3-flash` through the organizer's gateway.
- the text model of GLM-5.3-Flash (glm5_next) and GLM-5.3 (glm_moe_dsa),
models` and shows a red check. gt4py uses `zai-org/GLM-5.3-Flash` with fallbacks `Qwen/Qwen3.8-Flash-Next-FP8,Qwen/Qwen3.8-27B-FP8`.
## Verified against the real endpoint (`deepseek-v4.1-flash` / `glm-5.3-flash`, 21 models)
feat(deepinfra): add GLM-5.3-Flash, MiMo-V2.6-Flash, Nemotron-3.5-Super-VL, and Hy4-preview
M3 Ultra 256 GB, GLM-5.3-Flash, `msb-throughput --engine scheduler --prompt-tokens 1024 --decode-tokens 256 --runs 3`, the provider slice at `3675596d1` against main at `30ea73b89`, interleaved, two
paid model; the cheap tier (deepseek-v4.1-flash, glm-5.3-flash, qwen3.8-flash,
GLM-5.3 Flash ist ein Modell von Z.ai mit offenen Modellgewichten, veröffentlicht am 2026-08-26. Das Kontextfenster umfasst 1.048.576 Tokens. Die maximale Ausgabelänge beträgt 131.072 Tokens. Pro Million Tokens kostet die Eingabe $0,15 und die Ausgabe $0,50.
Das Kontextfenster umfasst 1.048.576 Tokens. Die maximale Ausgabelänge beträgt 131.072 Tokens.
Pro Million Tokens kostet die Eingabe $0,15 und die Ausgabe $0,50. Cache-Eingabe kostet $0,03 pro Million.
Ja. Z.ai veröffentlicht die Modellgewichte.
Am 2026-08-26.