Expired50% confidenceFactExact time
The 14B MoE model retains 95.9% of coding capability, with a loss of only approximately 4%
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026(expired)
Sources
Related Claims
UnverifiedDevelopers ultimately selected the variant with the lowest KL divergence (0.028), resulting in a 1.83% drop on GSM8K, no change on MMLU, and an improvement in HarmBench compliance rate from 30.8% to 100%65% similarUnverifiedThe 14B MoE model is derived by compressing the Qwen3.6 35B MoE, reducing total parameters from 35B to 14B, while the active parameters per forward pass remain at 3B, fully consistent with the original 35B62% similarUnverifiedThe 14B MoE model is a jailbroken version, with a KLD (divergence metric from the original model) of only 0.0015 (approximately 0.15%)61% similarUnverifiedOmniMerge V4 achieved 73.4% on the MBPP code test, an improvement of 15.8 percentage points compared to the original Q6K quantization's 56.2%.61% similarUnverifiedLing-3.0-flash adopts a Mixture of Experts (MoE) architecture with an activation ratio of only approximately 4% (approximately 5.1B activated out of 124B total parameters).60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/40893API
curl https://kongchang.com/api/v1/knowledge/claims/40893MCP
get_claim(id=40893)