Expired50% confidenceEventExact time
A Reddit user switched to the 3.8-27B model and discontinued the previously used 3.5/3.6-35B series models.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
9/12/2026
First Seen
Valid until: 9/26/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedThe 3.8-27B has fewer parameters than the 3.5/3.6-35B-A3B series.62% similarUnverifiedThe author noted that other models previously tested, such as 35B and 3B variants, actually performed worse than the Q2 version of Qwen3 27B on the Kanban task.60% similarUnverifiedIn tests on 2,456 held-out prompts, increasing the sampling steps from 20 to 50 improved FID only from 27.0 to 26.6 (a gain of 0.4), while removing the shift caused FID to rise from 27.0 to 27.3 and FD-DINOv2 to rise from 218 to 228.57% similarUnverifiedQwen2.5-3B-Instruct, Phi-3.5-mini (3.8B), and Llama-3.2-3B-Instruct are the three models with the best overall performance at small parameter scales at the end of 2024.55% similarUnverifiedThe user ultimately settled on a 27B uncensored model, having previously iterated through 27B and 35B MoE models in the Qwen series.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916967API
curl https://kongchang.com/api/v1/knowledge/claims/916967MCP
get_claim(id=916967)