GLM5 Autonomously Builds Working Game Boy Emulator

GLM-5 — What It Is

Developer: Z.ai (formerly Zhipu AI / ZhipuAI), a Tsinghua University spinoff founded in 2019. The company rebranded to Z.ai in 2025 and completed a Hong Kong IPO on January 8, 2026, raising approximately HKD 4.35 billion (USD $558 million).

• Official release date: February 11, 2026
• GitHub: https://github.com/zai-org/GLM-5
• HuggingFace weights: https://huggingface.co/zai-org/GLM-5
• API: https://api.z.ai
• OpenRouter: https://openrouter.ai/z-ai/glm-5
• License: MIT (fully open-source, commercial use permitted)
• Parameters: 744–745B total, ~40–44B active per token (MoE)
• Context window: 200K input tokens, up to 128K output tokens
• Pre-training data: 28.5 trillion tokens
• Pricing: ~$1.00/M input tokens, ~$3.20/M output tokens (approximately 5–8x cheaper than Claude Opus 4.5)

The company name in the transcript is slightly garbled — it is Z.ai (stylized), not "ZAI." The model family is the GLM (General Language Model) series, now in its fifth generation.

How It Works — Technical Architecture

GLM-5 uses a Mixture-of-Experts (MoE) architecture with 744B total parameters and approximately 40–44B active parameters per token, representing roughly a 2x scale-up from its predecessor GLM-4.5 (355B total, 32B active). Pre-training data grew from 23T to 28.5T tokens.

Key architectural innovations:
- DeepSeek Sparse Attention (DSA): Borrowed from DeepSeek, this mechanism enables efficient long-context handling up to 200K tokens without the typical quadratic memory cost.
- Slime async RL framework: A novel asynchronous reinforcement learning infrastructure developed by Zhipu AI. Traditional RL training at this scale is notoriously inefficient; Slime enables post-training runs producing 3,000–6,000 messages per run (~60–100M output tokens), specifically honing long-range planning and tool use.
- MoE routing: 78 transformer layers, 256 experts per layer with 8 activated (~5.9% sparsity). The first three layers are dense; subsequent layers use sparse attention.
- Hardware independence: GLM-5 was trained entirely on 100,000 Huawei Ascend 910B chips using the MindSpore framework — zero dependency on NVIDIA hardware. This is both a technical milestone and a geopolitical statement, as Zhipu AI has been on the U.S. Entity List since January 2025, blocking access to H100/H200 GPUs.

Benchmark performance (self-reported):
- SWE-bench Verified: 77.8% (#1 open-source; Claude Opus 4.5 leads at 80.9%)
- AIME 2026: 92.7% (matches Claude Opus 4.5 at 93.3%)
- GPQA-Diamond: 86.0% (vs Claude Opus 4.5 at 87.0%)
- Humanity's Last Exam (with tools): 50.4% (beats Claude Opus 4.5 at 43.4%)
- Terminal-Bench 2.0: 56.2–60.7
- Vending Bench 2: $4,432 simulated balance (#1 open-source)

Known limitations:
- Inference speed: ~17–19 tok/s, noticeably slower than NVIDIA-backed competitors (~25+ tok/s)
- Self-hosting requires ~8× A100 80GB GPUs (substantial infrastructure)
- Fewer third-party integrations vs. OpenAI/Anthropic ecosystems
- Knowledge cutoff is not officially published

The E01 Research GBA Emulator Experiment — What Actually Happened

E01 Research (e01.ai) gained early access to GLM-5 to stress-test its long-task capabilities. They designed what they called the "Emulator Challenge": build a Game Boy Advance emulator from scratch in JavaScript, embedded in a 3D rendered scene, using a single agent with no parallelism.

There were two distinct test conditions — and the video conflates them:

1. "Easy mode" (with reference code): E01 gave GLM-5 the gbajs open-source GBA emulator source code as reference. GLM-5 read the architecture, understood the design, and reimplemented it independently. Result: working core emulator, ROM loading, 3D scene. Live demo: https://e01.ai/gba

2. "Zero reference" mode (no code, no web search): Only a system prompt and a GBA hardware specification document. This ran 24+ hours straight. Result: CPU instruction set core was completed, but the full emulator was still in progress — it was NOT a finished, playable emulator.

The working demo shown in the video (the one where ROMs load and games play) is from the "easy mode" run with reference code, not the pure 24-hour autonomous run. The 24-hour zero-reference run produced a partial result.

How the agent maintained state across context resets: The prompt defined a meta-loop (work → test → log → advance), persisting state in files (/notes/progress.md, /notes/decisions.md, /notes/blockers.md). GLM-5 made 700+ tool calls and 800+ context switches without degradation — prior-generation models failed by looping, forgetting goals, or halting on erroneous tool calls.

A notable human intervention: During the experiment, GLM-5 got stuck trying to generate a 3D model of the GBA console from scratch (a task better suited to a human sourcing an asset). A human intervened to unblock it. E01 Research noted that setting explicit "pause-and-ask" thresholds is important for long-running agents.

Try It Yourself

Access GLM-5 via API

pip install z-ai-sdk

from z_ai_sdk import ZAI

client = ZAI(api_key="YOUR_API_KEY")  # Get key at api.z.ai

response = client.chat.completions.create(
    model="glm-5",
    messages=[
        {"role": "user", "content": "Write a Python function to parse a GBA ROM header."}
    ]
)
print(response.choices[0].message.content)

Self-host with vLLM (requires 8× A100 80GB)

# FP8 quantized version
vllm serve zai-org/GLM-5-FP8 \
  --tensor-parallel-size 8 \
  --gpu-memory-utilization 0.85 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 1 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --enable-auto-tool-choice \
  --served-model-name glm-5-fp8

Self-host with SGLang (recommended for Hopper/Blackwell GPUs)

docker pull lmsysorg/sglang:glm5-hopper  # For H100/H200
docker pull lmsysorg/sglang:glm5-blackwell  # For B100/B200

Try the live GBA emulator demo

Visit: https://e01.ai/gba (the "easy mode" result with reference code)

What The Creator Didn't Mention

1. The company name is wrong. The creator says "ZAI" — the correct name is Z.ai (formerly Zhipu AI). This matters for finding documentation, the API, and the GitHub org (zai-org).

2. The 24-hour run did NOT produce a working emulator. The video shows the "easy mode" result where GLM-5 was given the gbajs reference source code. The pure zero-reference 24-hour run only completed the CPU instruction set — the full emulator was still in progress. This is a significant omission that changes the impressiveness of the claim.

3. Human intervention was required. E01 Research documented that GLM-5 got stuck trying to generate a 3D GBA console model from scratch and a human had to intervene to unblock it. It was not fully autonomous.

4. GLM-5 was soft-launched as "Pony Alpha" on OpenRouter in early February 2026 before the official February 11 release — a stealth stress test that the AI community identified through benchmark analysis and GitHub PRs.

5. Geopolitical significance: GLM-5 was trained entirely on Huawei Ascend chips with no NVIDIA hardware — a direct response to U.S. export controls that placed Zhipu AI on the Entity List in January 2025.

6. Alternatives for long-horizon agentic coding: Qwen 3.5 (Alibaba, 397B MoE), DeepSeek-V3.2-Thinking, Kimi K2.5, and MiniMax M2.2 all launched in the same February 2026 "Spring Festival offensive" wave. Claude Opus 4.5 still leads on SWE-bench Verified (80.9% vs GLM-5's 77.8%).

7. Self-hosting is not consumer-grade. Running GLM-5 locally requires approximately 8× A100 80GB GPUs. Most users will need to use the API or OpenRouter.

8. The E01 Research blog post is the primary source — it is publicly available at https://blog.e01.ai/glm5-gameboy-and-long-task-era-64db7074a026 and contains the full technical methodology.