Jan-v2-VL-max-FP8-GGUF

Jan-v2-VL-max-FP8-GGUF

Jan-v2-VL-max-FP8 from janhq is a 30B-parameter vision-language model extending the Jan-v2-VL family, fine-tuned from Qwen3-VL-30B-A3B-Thinking using LoRA-based RLVR to enhance long-horizon execution stability with minimal error accumulation over many steps, evaluated on "The Illusion of Diminishing Returns" benchmark emphasizing execution length over recall. Optimized for agentic automation and UI control tasks where plans/knowledge are provided upfront—such as stepwise browser/desktop operations with screenshot grounding and Jan Browser MCP tool calls—it shows no regressions and small gains versus its base on standard tasks, with largest improvements in extended sequences, while FP8 quantization reduces memory/latency for vLLM 0.12.0 deployment (transformers==4.57.1, llm-compressor) using parameters like temperature=1.0, top_p=0.95, presence_penalty=1.5. Hosted on Jan Web (chat.jan.ai) for immediate use, it supports --enable-auto-tool-choice and Hermes/DeepSeek_R1 parsers for production-ready multimodal agents in long-context scenarios.

Jan-v2-VL-max-FP8 [GGUF]

Quants Usage

(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)

Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):