Tev-320B

🌐 Project Page  |  📄 Paper (arXiv)  |  🤗 Tev-27B

Tev-320B is a long-horizon terminal agent model trained from GLM-5.3-Flash (320B total / 18B active parameters, mixture-of-experts) with the Tev recipe from the paper Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards. The recipe is reinforcement learning on Tev-Gym environments with progress-potential reward shaping (PPRS), the same recipe that produces Tev-27B.

Highlights

• Strongest open model on our Terminal-Bench "CM" subsets. Pass@1 rises from 0.20 to 0.27 (+35%) on TB 3.0-CM and from 0.29 to 0.37 (+28%) on TB 4.0-CM over GLM-5.3-Flash.
• On TB 4.0-CM, it trails only Claude Opus 5, Claude Fable 5, and GPT-5.6 (sol) in Pass@1, ahead of Grok 4.6, Claude Opus 4.8, and Claude Sonnet 5. On TB 3.0-CM it edges past Grok 4.5 and Claude Opus 4.8.
• Gains in every category on both suites. The largest is on TB 3.0 security, where Pass@1 rises from 0.22 to 0.45.

How Tev works

Open models struggle on long-horizon tasks for two reasons: there is almost no long-horizon training data, and terminal-only rewards are too sparse to learn from. Tev addresses both.

Tev-Gym: latent-world-first environment synthesis. Each task is built backward from a latent world, a complete, correct, hidden specification of the task domain, its workflow (a DAG of stages), and the expected state after every step. Agent-facing inputs are rendered from that world and corrupted with controlled, invertible defects that the agent must repair. Because the correct answer exists before the task, the latent world serves as an exact answer key. It grades the final submission and also every intermediate state. Tev-Gym contains 1,170 verified long-horizon environments across 7 domains (software engineering, data engineering, AI/CS research, mathematics and operations research, security and forensics, infrastructure administration, and quantum computing).

PPRS: dense progress rewards. After every turn, the agent's workspace is scored against the latent world's known-correct intermediate states, giving a progress potential Φ(s) ∈ [0, 1]. The rollout reward is

R(τ) = 1[τ passes the strict verifier] + λ · Φ(s_T),   λ = 0.25

applied with per-turn credit (return-to-go, γ = 1) and a per-turn-index group baseline. Φ separates rollouts that all fail by how far each one got, so groups that would carry no signal under terminal-only reward still train the policy.

Model details

Base model	zai-org/GLM-5.3-Flash
Parameters	320B total, 18B active (mixture-of-experts: 288 routed experts, 8 per token, plus 1 shared)
Training	RL with DPPO + PPRS on Tev-Gym (the Tev recipe)
Agent setting	Single bash tool in a per-task container
Weights	bf16, 49 safetensors shards (~643 GB). The base model's vision encoder is included unchanged. RL training and our evaluations are text-only. No MTP weights.
License	CC BY-NC 4.0, non-commercial use only

Evaluation

Terminal-Bench 3.0 and 4.0, "CM" subsets

The computing-and-math subsets contain 22 TB 3.0 tasks and the 18 of them that remain in TB 4.0. Leaderboard rows are from the official per-trial results (5 trials per task). We evaluated GLM-5.3-Flash, Tev-320B, and GLM 5.2 on TB 4.0 ourselves (GLM 5.2 has no TB 4.0 leaderboard entry). On the leaderboard models, subset scores track full-benchmark scores closely (Pearson r = 0.96 for TB 3.0 P@1 and 0.97 for TB 4.0 P@1).

Model	TB 3.0 P@1	TB 3.0 P@5	TB 4.0 P@1	TB 4.0 P@5
Claude Opus 5	0.53	0.64	0.52	0.72
Claude Fable 5	0.41	0.64	0.46	0.61
GPT-5.6 (sol)	0.41	0.55	0.39	0.61
Grok 4.6	0.30	0.50	0.31	0.50
Claude Opus 4.8	0.19	0.50	0.27	0.50
Grok 4.5	0.26	0.55	0.20	0.33
Claude Sonnet 5	0.16	0.32	0.19	0.50
GLM 5.2	0.04	0.18	0.03	0.13
GLM-5.3-Flash (base)	0.20	0.50	0.29	0.56
Tev-320B (ours)	0.27	0.55	0.37	0.56

Per category (P@1, base → Tev-320B):

Category	TB 3.0-CM	TB 4.0-CM
Software	0.18 → 0.21	0.16 → 0.25
ML	0.35 → 0.40	0.60 → 0.65
Operations	0.05 → 0.10	0.05 → 0.15
Security	0.22 → 0.45	0.68 → 0.70
All	0.20 → 0.27	0.29 → 0.37

Most of the gain is in P@1. P@5 rises by 0.05 on TB 3.0 and is unchanged on TB 4.0, so training mainly makes the model solve the tasks it can already reach more reliably, rather than unlocking new ones. See the paper for the full tables and evaluation protocol.

Usage

Tev-320B is a terminal agent. It works best in a loop that gives it a single bash tool and executes its commands in a sandbox, the setting it was trained in. The model loads with the same inference stacks as GLM-5.3-Flash. See the base model's SGLang cookbook and vLLM recipes.

Serve with SGLang

This is the configuration we used for our evaluations (SGLang v0.5.20, tensor parallelism over 4 B300 GPUs):

python3 -m sglang.launch_server \
  --model-path gordonhu/Tev-320B \
  --served-model-name tev-320b \
  --tp-size 4 \
  --context-length 262144 \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --language-only \
  --trust-remote-code \
  --mem-fraction-static 0.85

Call it as an agent

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")

tools = [{
    "type": "function",
    "function": {
        "name": "bash",
        "description": "Execute a bash command. Each command runs in a new subshell.",
        "parameters": {
            "type": "object",
            "properties": {"command": {"type": "string", "description": "The bash command to execute."}},
            "required": ["command"],
        },
    },
}]

messages = [
    {"role": "system", "content": (
        "You are a helpful coding assistant. You have access to a bash terminal. "
        "When you are confident your solution is correct, submit by running: "
        "`echo COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT`"
    )},
    {"role": "user", "content": "Read /task/prompt.md and complete the task described there."},
]

response = client.chat.completions.create(
    model="tev-320b", messages=messages, tools=tools,
    temperature=1.0, top_p=0.95,
)
print(response.choices[0].message.tool_calls)
# Run each `bash` call in your sandbox, append the output as a "tool" message, and repeat
# until the model submits or your turn budget runs out.

Intended use and limitations

• Research use only. Tev-320B is released for non-commercial research on long-horizon agents (see License).
• Run it in a sandbox. The model executes shell commands autonomously. Run it in an isolated container with no access to credentials, production systems, or data you cannot afford to lose.
• Gap to frontier models. Tev-320B still trails Claude Opus 5, Claude Fable 5, and GPT-5.6 (sol) on TB 4.0-CM.
• Evaluation scope. The paper evaluates Tev-320B on the TB 3.0 and 4.0 "CM" subsets only. Training covers terminal-based computing-and-math workflows in synthetic environments. Multimodal abilities were neither trained nor evaluated.
• Dual use. Like any more capable coding agent, the model can be misused. The security-flavored tasks in Tev-Gym are analysis and forensics exercises over synthetic artifacts.

License

Tev-320B is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. You may use, share, and adapt the model for non-commercial purposes only, with attribution. Commercial use is not permitted.

Tev-320B is a derivative of GLM-5.3-Flash, Copyright (c) 2026 Z.AI Co., Ltd., which is licensed under the MIT License. The original copyright and permission notice is included as LICENSE-GLM-5.3-Flash.

Citation

@misc{hu2026tev,
  title  = {{Tev}: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards},
  author = {Hu, Wenbo and Pramanick, Shraman and Singh, Krishna Kumar and
            Peng, Nanyun and Chang, Kai-Wei and Lee, Yong Jae},
  year   = {2026},
  eprint = {2610.18868},
  archivePrefix = {arXiv}
}

Notice

The model weights held in this repo are a reproduced version, not official model weights trained at Adobe.