Qwen3-14B
An open-source language model in Alibaba's Qwen3 series with 14 billion parameters, featuring capabilities in reasoning and instruction following.
Core Facts
Timeline (last 90 days)
Carbon-X is built on Qwen3-14B.
Using Qwen3-14B to generate risk exam papers yielded only 5/25 passages with clear signals, and only after switching to stricter prompts and Qwen3-32B were 21 passages obtained.
There is a case on Reddit: a team using Qwen3-14B running on vLLM found that in its LangGraph multi-agent system, the monthly Token consumption of just the single Supervisor node equaled the sum of all 8 of its expert Agents.
Researchers tested two technical approaches on the Qwen3-14B model: directly training the model to learn when to answer, and training an independent answer selector
The reinforcement learning modifications tested in the study did not yield consistent gains, and directly training models to learn to abstain proved unstable.
The large disparity in error reduction between Gemma-2-9B and Qwen3-14B indicates that different architectures differ significantly in their sensitivity to task context.
On Gemma-2-9B and Qwen3-14B, the researchers successfully switched the facts read by the model by modifying early internal activation vectors while keeping the written instructions unchanged.
The base model is Qwen3-14B, from which two versions—KGModel and CGModel—are trained separately via supervised fine-tuning (SFT).
The off-policy distillation experiment uses Qwen3 0.6B as the student model and Qwen3 4B as the teacher model.
All Facts (8)
The off-policy distillation experiment uses Qwen3 0.6B as the student model and Qwen3 4B as the teacher model.
75%UnverifiedCarbon-X is built on Qwen3-14B.
50%UnverifiedUsing Qwen3-14B to generate risk exam papers yielded only 5/25 passages with clear signals, and only after switching to stricter prompts and Qwen3-32B were 21 passages obtained.
50%UnverifiedResearchers tested two technical approaches on the Qwen3-14B model: directly training the model to learn when to answer, and training an independent answer selector
50%UnverifiedThe reinforcement learning modifications tested in the study did not yield consistent gains, and directly training models to learn to abstain proved unstable.
50%UnverifiedOn Gemma-2-9B and Qwen3-14B, the researchers successfully switched the facts read by the model by modifying early internal activation vectors while keeping the written instructions unchanged.
50%UnverifiedThe large disparity in error reduction between Gemma-2-9B and Qwen3-14B indicates that different architectures differ significantly in their sensitivity to task context.
50%UnverifiedThe base model is Qwen3-14B, from which two versions—KGModel and CGModel—are trained separately via supervised fine-tuning (SFT).
50%