Jan-v2-VL-max extends the Jan-v2-VL family to a
30B-parameter vision–language model focused on
long-horizon execution. This release scales model capacity and applies
LoRA-based RLVR to improve stability over many steps with
low error accumulation. For evaluation, we continue to use
The Illusion of Diminishing Returns: Measuring Long-Horizon Execution in LLMs, which emphasizes execution length rather than knowledge recall.
Tasks where the plan and/or knowledge can be provided up front, and success hinges on stable, many-step execution with minimal drift:
Evaluated under FP8 inference,
Jan-v2-VL-max vs.
Qwen3-VL-30B-A3B-Thinking shows
no regressions and
small gains on several tasks, with the largest improvements in
long-horizon execution. Our FP8 build maintains accuracy while reducing memory footprint and latency.
Hosted on
Jan Web — use the model directly at
chat.jan.ai1# Exact versions used in our evals
2pip install vllm==0.12.0
3pip install transformers==4.57.1
4pip install "git+https://github.com/vllm-project/llm-compressor.git@1abfd9eb34a2941e82f47cbd595f1aab90280c80"
1vllm serve Menlo/Jan-v2-VL-max-FP8 \
2 --host 0.0.0.0 \
3 --port 1234 \
4 -dp 1 \
5 --enable-auto-tool-choice \
6 --tool-call-parser hermes \
7 --reasoning-parser deepseek_r1
8
For optimal performance in agentic and general tasks, we recommend the following inference parameters:
1temperature: 1.0
2top_p: 0.95
3top_k: 20
4repetition_penalty: 1.0
5presence_penalty: 1.5
