Gravity 30B-A5B RL
A 30B mixture-of-experts model (5B active, 64 experts top-8, MLA attention) tuned for multi-turn tool use.
Results
BFCL v3, three runs per checkpoint on fresh servers, mean ± standard deviation.
| Live Acc (n=1351) | multi-turn (n=800) | |
|---|---|---|
| SFT base | 69.97 ± 0.99 | 27.05 ± 0.29 |
| this model | 73.11 ± 0.34 | 30.25 ± 0.50 |
LAB-Bench (six non-image subtasks, tools on) shows no measurable change: 48.24 for the base against 49.58 here, against a baseline that spans 6.4 points across its own two runs.
Serving
The 131,072-token window is required — BFCL's long-context problems overflow a shorter one and score zero.
vllm serve trillionlabs/Gravity-30BA5B-RL \
--tensor-parallel-size 8 --max-model-len 131072 --trust-remote-code \
--enable-auto-tool-choice --tool-call-parser cyan_glm \
--tool-parser-plugin <kit>/serve/glm_tool_parser.py --reasoning-parser glm45
Without the reasoning parser the model's <think> block lands in content and answers stop
parsing. Both plugins ship with the evaluation kit.
Reproducing
The harness that produced the table is gravity-bfcl-labbench. One command:
./gravity-bench --model trillionlabs/Gravity-30BA5B-RL