Unverified50% confidenceBenchmarkExact time
在文档摘要基准测试上，Qwen2.5-32B-Instruct和Llama-3.1-70B的量化版本已能达到GPT-4早期版本85%-92%的表现水平
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
8/10/2026
First Seen
Valid until: 11/8/2026
Sources
Related Claims
UnverifiedOpen-source models such as Qwen-7B and LLaMA-2-7B, after Agent Tuning, can achieve tool-calling accuracy at the level of GPT-3.5 or even close to GPT-480% similarUnverifiedThe 4-bit quantized version of Qwen3-32B has achieved performance on tasks such as code generation and Chinese language understanding that is close to early GPT-4 levels.80% similarUnverifiedTop autoregressive models such as GPT-4o and Claude 3.5 have already exceeded 95% accuracy on GSM8K, leading to benchmark saturation.75% similarUnverifiedQwen2.5-Coder demonstrates performance approaching GPT-4o on code generation benchmarks such as HumanEval, and its 7B version can run on an ordinary laptop after quantization.74% similarVerifiedMeta's released Llama 3.1 405B has approached GPT-4 levels across multiple benchmark tests.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/723535API
curl https://kongchang.com/api/v1/knowledge/claims/723535MCP
get_claim(id=723535)