Tev1-4B-experimental-GGUF

Tev1-4B-experimental-GGUF

Tev1-4B-experimental is an experimental 4-billion-parameter decision model from Together AI, a supervised fine-tune of Qwen3.5-4B trained to select a single option from a structured state, question, and list of 2–24 labeled choices — a "Jev-inspired" experiment that retains Qwen's standard autoregressive next-token head rather than implementing a true non-autoregressive Jev runtime. Its intended interface takes a system instruction plus a JSON-structured decision payload (treating the state field strictly as data rather than instructions) and returns just a single option letter at temperature=0 with enable_thinking=False, making it usable via the Together API or compatible chat completion endpoints. On Together's internal development evaluation, it scored 88.0% (880/1,000) on the main decision set and 100% (300/300) on a synthetic policy-transfer set, with all outputs valid single letters — though the authors caution these are development results without an untuned-Qwen baseline for comparison, not an independent benchmark. It is explicitly not intended for generic chat use, should not serve as the sole authority for high-impact decisions, and has not been comprehensively evaluated for prompt injection resistance, multilingual behavior, or calibration; the base Qwen3.5-4B is Apache-2.0, but the release license for these fine-tuned weights was still being finalized at time of publication.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp