Unverified50% confidenceFactExact time
Model screening is performed automatically by survey.py, the license must be Apache-2.0 or MIT, and at least 2 GiB must be left for the KV cache; in the end Tier 1 has 8 models (native JAX) and Tier 2 has 7 models (PyTorch fallback torchax).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
单块TPU v5e实测：哪些小模型能与Gemma 4 E2B一战？
Related Entities
Related Claims
UnverifiedNumerical consistency testing can be systematized into a regression test suite that computes the maximum absolute error and KL divergence of per-token log probabilities from both sides after each engine version upgrade, with a tolerance threshold (e.g., KL < 1e-4) that blocks training when exceeded.55% similarUnverifiedGPTQ、AWQ等后训练量化方法能够在几乎不损失基准测试分数的情况下将70B参数量级模型部署在消费级GPU上55% similarUnverifiedSWE-bench Verified's scoring mechanism relies on an automated test suite; after a model submits a code patch, the system runs predefined unit tests to determine whether they pass55% similarUnverifiedWhen evaluating whether a model can be deployed locally, one should pay attention to the native parameter scale, weight format, availability of quantized versions, recommended minimum hardware configuration, and open-source license terms.55% similarUnverifiedSWE-bench测试样本取自真实GitHub仓库的历史Issue，评分标准是能否通过仓库原有的完整测试套件54% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1014585API
curl https://kongchang.com/api/v1/knowledge/claims/1014585MCP
get_claim(id=1014585)