Unverified50% confidenceBenchmarkExact time
The author tested Qwen3 Flash Next using dual 7900 XTX graphics cards, Qwen3 Q3_K_XL quantization, and a 32K context, achieving capability scores of 88 in knowledge reasoning, 100% in instruction following, 30% in advanced mathematics, 90 in Chinese, and 80 in code.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/17/2026
First Seen
Valid until: 12/16/2026
Sources
Related Entities
Related Claims
UnverifiedQwen3.8 Flash Next has 125 billion total parameters, with only approximately 6 billion parameters activated per token during inference70% similarUnverifiedThe author plans to use the same set of tools to test Qwen3 27B and make a direct comparison with Flash Next.69% similarUnverifiedOn the comprehensive intelligence index AA, Qwen3.8 Flash Next scored 56, the cloud version of Qwen3.8 Flash scored 57, and DeepSeek V4 Flash scored 52.68% similarUnverifiedQwen3 27B achieved a 100% pass rate in the 256K long-context memory retrieval test (needle in a haystack), with all 15 runs being accurate.66% similarUnverifiedQwen Coder adopts a Mixture of Experts (MoE) architecture with 480 billion total parameters, activating only 35 billion parameters during inference.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930283API
curl https://kongchang.com/api/v1/knowledge/claims/930283MCP
get_claim(id=930283)