# glm-5.3-flash decision evidence

GLM-5.3-Flash is Z.ai's MIT-licensed open-weight mixture-of-experts model (320B total, 18B active) with a 1M-token context window and video, image, text and file input. Thinking cannot be turned off; reviewed tests show strong single results but uneven completion across tasks.

Verified: 2026-09-24; snapshot: 2026-09-25.

## Published model card

As the vendor publishes it for the model and its weights; these are not the limits of a hosted endpoint.

- Context window: 1M tokens (1,048,576)
- Maximum output: 128K tokens (131,072 on the Z.ai API)
- Modalities: Video, image, text and file input; text output
- Parameter count: 320B total, 18B active (MoE)

## Open weights and deployment

Status: Available. Official FP8 and BF16 safetensors are published under the MIT License. The publisher lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth for local serving.

### Agent deployment prompt

You are my deployment agent. Deploy the official zai-org/GLM-5.3-Flash weights as a private, OpenAI-compatible service on this host. Use the publisher's model card as the source of truth: https://huggingface.co/zai-org/GLM-5.3-Flash.

Preflight
- Inspect the OS, GPU model and count, free VRAM, driver and CUDA versions, free disk space, and available ports. The default repository holds block-wise FP8 weights; zai-org/GLM-5.3-Flash-BF16 is the official BF16 alternative. Estimate whether the chosen weights and the requested context fit with safe headroom for runtime and concurrency; Z.ai publishes no minimum GPU requirement.
- If the host cannot safely run the official weights, stop before installing or launching anything. Explain the limiting resource and propose a separately reviewed alternative; do not substitute a community quantization automatically.

Deployment
- If preflight passes, create an isolated environment, install a current SGLang or vLLM release that documents GLM-5.3-Flash support, record the exact model revision, and serve the FP8 repository through its OpenAI-compatible API. Generate an API key and store it in a protected environment variable. Bind to 127.0.0.1 first; do not expose the endpoint publicly.
- Keep the publisher's sampling defaults (temperature 1.0, top_p 0.95). Thinking is always on for this model, so do not try to disable it. Start with a context length the preflight supports; the model supports up to 1,048,576 tokens.

Verification and handoff
- Send a real request to /v1/chat/completions and confirm that the model responds. Report the exact launch command, installed versions, model revision, local endpoint, test result, measured GPU memory and log location. Identify any reverse-proxy or firewall changes that still need my approval. Never print the API key or other secrets.

- Repository: https://huggingface.co/zai-org/GLM-5.3-Flash
- Licence: MIT
- Recommended runtime: SGLang or vLLM
- Hardware note: Not published by Z.ai; the Agent recipe performs a hardware preflight before serving.

## Real-task reports

### [RemakeBench Launch 005: four Flash models on 14 fixtures](https://www.remakebench.com/qwen-3-8-flash-vs-deepseek-v4-1-flash-vs-glm-5-3-flash-vs-gemini-3-8-flash)

By [RemakeBench](https://www.youtube.com/@remakebench) (@remakebench). Benchmark channel that publishes per-fixture clips, costs, wall-clock times and validation notes

Published: 2026-09-18; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Fourteen fixtures spanning game building, 3D scenes, CAD, drawing and robot control, run on Qwen 3.8 Flash, DeepSeek V4.1 Flash, GLM 5.3 Flash and Gemini 3.8 Flash
- Method: Each model's attempt per fixture is shown as a video clip with its displayed cost, wall-clock time and reasoning setting; pass/fail labels are the lab's video judgments.
- Finding: No model won overall. GLM 5.3 Flash was the only model to pass the Apple-stem manipulation task and passed the single-turn Rubik's Cube, but did not complete the three-move extension and delivered no native geometry in Turbofan CAD.
- Limitation: Pass/fail labels are video judgments without a machine-readable validator, and single attempts do not establish reliability; the lab publishes no aggregate score or overall winner. GLM's Infinite Cathedral completion needed repeated user continuations because the endpoint stalled.
- Apple-stem manipulation: Pass · $0.5583 · 1h 29m 33s
- Rubik's Cube single turn: Pass · $0.3414 · 40m 48s

### [GLM 5.3 Flash alone and with Jev on Halite 2](https://x.com/Sentdex/status/2101828851458293827)

By [Harrison Kinsley](https://x.com/Sentdex) (@Sentdex). Developer who ran this Halite 2 experiment and answered questions about it in the thread

Published: 2026-09-21; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Play the turn-based strategy game Halite 2
- Method: Compared GLM 5.3 Flash alone with a hybrid in which GLM makes the strategic decisions and Jev executes ship-by-ship moves, across seeded maps; in the thread the author cites 133 unique maps.
- Finding: GLM 5.3 Flash alone beat Jev alone 82% of the time; the hybrid was 13× faster at 56% of the pure GLM API cost with a slight performance edge.
- Limitation: The author calls it a single example and among their first experiments with Jev; results may not carry over to other tasks.
- GLM 5.3 Flash vs Jev win rate: 82%
- Hybrid speed-up: 13×
- Hybrid API cost: 56% of pure GLM

### [DeepSeek V4.1 Flash and GLM 5.3 Flash recreating two game menus](https://x.com/MiaAI_lab/status/2100134727457902942)

By [Mia](https://x.com/MiaAI_lab) (@MiaAI_lab). X account that posts LLM experiments and attached the full output videos

Published: 2026-09-16; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Recreate the Psychonauts 2 and Metroid Prime main menus as closely as possible to the originals
- Method: One shot per game, both models at Max thinking in the same harness; the author attached full-length videos and reported tokens used.
- Finding: GLM 5.3 Flash used 134k and 111k tokens, a third to a quarter of DeepSeek V4.1 Flash's, and the author judged its menus clearly better on both games.
- Limitation: Which menu is closer to the original is the author's own judgement on two prompts, without a scoring rubric.
- Tokens · Psychonauts 2: 134k
- Tokens · Metroid Prime: 111k

### [DeepSeek V4.1 Flash and GLM-5.3-Flash on a Blender geometry task, run locally](https://x.com/WescheNex1q/status/2100759267259117656)

By [Wësche](https://x.com/WescheNex1q) (@WescheNex1q). Artist and AI enthusiast on X who builds and benchmarks open models on DGX Spark clusters

Published: 2026-09-18; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Write a Blender script that builds a dodecahedron trapped inside a dodecahedron
- Method: One shot with thinking on, temperature 0.6, no token cap and no retries; the script ran headless in Blender 5.2 and the author checked the mesh independently. DeepSeek ran as FP8 on vLLM across four DGX Sparks, GLM as an EXL3 quantisation across two.
- Finding: GLM-5.3-Flash produced a correct mesh (zero intersections, containment verified in-script) and the nicer picture, but took 93 min 12 s and 132,250 tokens at 23.7 tok/s, against 20 min 15 s and 62,928 tokens for DeepSeek V4.1 Flash.
- Limitation: One local run per model on different hardware and precision: GLM ran as a third-party EXL3 quantisation on two DGX Sparks and DeepSeek as FP8 on four, so speed and token counts describe those builds, not the official weights or a hosted API.
- GLM-5.3-Flash: 93 min 12 s · 132,250 tokens
- DeepSeek V4.1 Flash: 20 min 15 s · 62,928 tokens

### [Three Flash models build the same koi pond locally](https://x.com/WescheNex1q/status/2102527793057661386)

By [Wësche](https://x.com/WescheNex1q) (@WescheNex1q). Artist and AI enthusiast on X who builds and benchmarks open models on DGX Spark clusters

Published: 2026-09-22; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: A koi pond build, run as a local head-to-head
- Method: The same task on GLM-5.3-Flash, MiMo-V2.6-Flash and DeepSeek V4.1 Flash, each on two DGX Sparks; the author reported time, tokens and fixes needed and left the visual verdict partly to readers.
- Finding: All three finished cleanly with zero fixes, but GLM-5.3-Flash was slowest at 59 minutes and 90.9K tokens, against 30 minutes for MiMo-V2.6-Flash and 21 for DeepSeek V4.1 Flash; the author said it was the first of their tests where GLM was not the winner.
- Limitation: One local run per model; the post states neither the prompt nor the quantisation or runtime used.
- Time: 59 min
- Tokens: 90.9K

### [AI Coding Daily LLM Coding Leaderboard](https://aicodingdaily.com/leaderboard)

By [AI Coding Daily](https://www.youtube.com/@AICodingDaily) (@AICodingDaily). Coding-model testing channel and website that runs its own leaderboard and sells premium tutorials and channel advertising

Published: 2026-09-24; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Implement and harden production-shaped projects in several stacks, including a Laravel app, a Go service and a React/TypeScript app, from the same prompts
- Method: Each model and reasoning setting is a separate row run in a coding harness, with the average cost and time per run. Code quality is graded by a judge model against a published 100-point rubric and scaled to 20 points per project; behavioural reliability runs each of four projects five times and scores the checks every attempt fails.
- Finding: GLM-5.3-Flash at max reasoning scored 42.82 ($0.02 and 8 min 54 s per run in OpenCode), 38th of 47 rows and well below GLM-5.3 at 49.92.
- Limitation: One author's benchmark whose code-quality half is graded by a judge model; rows run in different harnesses, the page is updated as models are added (read on 2026-09-24, 47 rows), and the author sells premium tutorials and advertising.
- Points (max): 42.82
- Rank: 38 of 47

### [kzhu's six-question LLM test leaderboard](https://www.techgogogo.com/llm-benchmark/)

By [AI产品狙击手 kzhu](https://www.youtube.com/@kevinzhu9305) (@kevinzhu9305). Chinese YouTube reviewer who scores models on a six-question test and publishes every score with per-question notes

Published: 2026-09-24; retrieved: 2026-09-24; verified: 2026-09-24.

- Task: Six questions scored out of 10: constrained sentence writing, two reasoning puzzles (replaced by pelican and compound-bow SVG animations in version 2.2), a browser OS with a space shooter, a 3D racing game and a Trello-style board
- Method: Each row names the harness the model ran in (Pi, Claude Code, WorkBuddy and others) and the date; the author scores every question with a note, and warns that totals are only comparable within one test version.
- Finding: GLM-5.3-Flash scored 54/60 on version 2.1 in zcode, with every reasoning answer right and the best recent space shooter the author had seen, but 45 after re-running the two version-2.2 SVG questions in Pi, where the pelican sank below the ring and the bow pointed backwards, taking over 20 minutes against under 10 for DeepSeek.
- Limitation: One author's scores on six questions; rows use different harnesses and test versions (2.1 or 2.2), and a few rows' table total differs by one point from the author's note.
- Score: 54 / 60 (v2.1, zcode)

## Public leaderboard coverage

A board missing from this list has no figure for this model in this snapshot; missing coverage is not a zero.

- LMArena Text Arena — Arena score (Elo): 1475 ±7 (retrieved 2026-09-22) [source](https://arena.ai/leaderboard/text)
- LiveBench — Overall: 71.6 (retrieved 2026-09-12) [source](https://livebench.ai/)
- Artificial Analysis Intelligence Index — Intelligence Index: 42 (retrieved 2026-09-21) [source](https://artificialanalysis.ai/leaderboards/models)
- SuperCLUE 智能指数 — 总分 (overall): 68.10 (retrieved 2026-09-22) [source](https://www.superclueai.com/homepage)
- Epoch Capabilities Index (ECI) — General ECI: 152 (150 - 154) (retrieved 2026-09-22) [source](https://epoch.ai/eci)
- Vals Index — Accuracy (GDP-weighted finance, coding and legal): 47.22 ±1.45 (retrieved 2026-09-22) [source](https://www.vals.ai/benchmarks/vals_index)
- LMArena Text Arena · Chinese — Arena score (Elo): 1529 ±25 (retrieved 2026-09-22) [source](https://arena.ai/leaderboard/text/chinese)
- LMArena Text Arena · Coding — Arena score (Elo): 1525 ±12 (retrieved 2026-09-22) [source](https://arena.ai/leaderboard/text/coding)
- LiveBench · Coding — Coding: 79.0 (retrieved 2026-09-12) [source](https://livebench.ai/)
- LiveBench · Agentic Coding — Agentic Coding: 56.8 (retrieved 2026-09-12) [source](https://livebench.ai/)
- Artificial Analysis · Output Speed — Median output tokens/s: 65 (retrieved 2026-09-22) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · GDPval-AA v2.1 — Agentic Real-World Work Tasks, (Elo-500)/2000 (%): 57 (retrieved 2026-09-21) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · Terminal-Bench v4.0 — Agentic Coding & Terminal Use (%): 33 (retrieved 2026-09-21) [source](https://artificialanalysis.ai/leaderboards/models)
- Artificial Analysis · MLCR-AA — Medical Long Context Reasoning (%): 51.1 (retrieved 2026-09-07) [source](https://artificialanalysis.ai/evaluations/mlcr-aa)

## Curated comparisons

- [deepseek-flash vs glm-5.3-flash](https://dash.chinaapi.ai/models/compare/deepseek-flash-vs-glm-5.3-flash)
- [gemini-3.8-flash vs glm-5.3-flash](https://dash.chinaapi.ai/models/compare/gemini-3.8-flash-vs-glm-5.3-flash)
- [glm-5.3-flash vs mimo-v2.6-flash](https://dash.chinaapi.ai/models/compare/glm-5.3-flash-vs-mimo-v2.6-flash)
- [glm-5.3-flash vs qwen3.8-flash](https://dash.chinaapi.ai/models/compare/glm-5.3-flash-vs-qwen3.8-flash)
- [glm-5.3 vs glm-5.3-flash](https://dash.chinaapi.ai/models/compare/glm-5.3-vs-glm-5.3-flash)

## Official sources

- [GLM-5.3-Flash: Frontier Intelligence, Flash Cost](https://z.ai/blog/glm-5.3-flash) — Z.ai, retrieved 2026-09-24
- [GLM-5.3-Flash model documentation](https://docs.z.ai/guides/vlm/glm-5.3-flash) — Z.ai, retrieved 2026-09-24
- [GLM-5.3-Flash model card](https://huggingface.co/zai-org/GLM-5.3-Flash) — Z.ai, retrieved 2026-09-24