Unverified50% confidenceBenchmarkExact time
Some models show a noticeable drop in needle-in-a-haystack accuracy beyond 32K tokens, while others remain stable over longer ranges.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/8/2026
First Seen
Valid until: 1/6/2027
Sources
Related Entities
Related Claims
UnverifiedIf samples exceeding 32K tokens rarely appear in a model's pre-training corpus, the model's performance in that range tends to be unstable even when the context window is extended through positional encoding.74% similarUnverified即便是支持128K上下文的模型，在文档长度超过32K时，中段信息的检索准确率往往出现明显下滑70% similarUnverifiedThe quality of long-context understanding in small models degrades rapidly with length, typically deteriorating after exceeding a 4K–8K token window.69% similarUnverifiedThe periodically refreshed model achieves higher accuracy and answer completion rates under short budgets, but when the output limit is relaxed to 32K, the frozen-response model's average accuracy surpasses it.68% similarUnverifiedThe 1-bit quantization of Qwen3 27B causes a cliff-like drop in model capability, with some task outputs losing basic coherence and accuracy.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/988422API
curl https://kongchang.com/api/v1/knowledge/claims/988422MCP
get_claim(id=988422)