Unverified50% confidenceOpinionExact time
This analysis is based on only 113 tasks from the DeepSWE benchmark; the sample size is limited, and the generalizability of the conclusions requires validation on a larger scale.
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified该测试样本规模相对有限（5个模型、13个代码库），结论有待更大规模验证74% similarUnverifiedA carefully curated and deduplicated high-quality dataset can train a model of equal or even superior capability using far fewer tokens than the original scale.65% similarUnverifiedFor structured tasks such as code completion and syntax checking, small models with 7B–13B parameters can achieve over 90% accuracy, with inference costs as low as 1/50 or less of those of top-tier models.64% similarUnverifiedOn a test subset of 152 completely untouched questions, anchored hybrid retrieval achieved 42.11%, while the planner-guided method reached only 36.84%.64% similarUnverifiedEven models that claim to support context windows of 128K or even million-level tokens frequently reveal weak mid-range retrieval in Needle-in-a-Haystack tests.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/946780API
curl https://kongchang.com/api/v1/knowledge/claims/946780MCP
get_claim(id=946780)