6 results on this page
A TraceML benchmark using Qwen3-1.7B with LoRA found that equivalent effective batch sizes can have substantially different training times depending on physical batch size and accumulation steps. Across 100 optimizer updates, 4×1 was 17% faster than 1×4 on a T4 and 41% faster on an L4.
Europe’s data centre capacity is projected to reach 13 gigawatts by the end of 2026, up roughly 20% year over year. In the second quarter, 63% of approximately 700 megawatts of new capacity went outside Frankfurt, London, Amsterdam, Paris, and Dublin.
The post argues that Qwen3.8 intermediate tokens should not be interpreted as human-like reasoning. Citing research, it highlights that trace validity often does not correlate with answer correctness, and that models trained on corrupted or irrelevant traces can perform as well as or better than models trained on valid traces.
Kimi K3 and Qwen3.8-Max have reportedly introduced revenue-sharing clauses into their open-source strategies. The shift suggests that open models may be moving from unrestricted free use toward a freemium model.
QwenMix-3.7 is an experimental merge of the Qwen3.8-27B and Qwen3.6-27B models. The creator reports that it passed a basic smoke test, but no meaningful evaluation has been performed yet.
Qwen Code v0.21.14 successfully completed one SWE-bench Verified task and one Terminal-Bench 2.0 task. Both benchmarks recorded 100% scores with no unresolved cases, execution errors, or infrastructure failures.