Latent Sparse Attention (LSA) is released as a step toward subquadratic attention that matches full-attention accuracy on ultra-long contexts. It makes sparse attention up to 22.7× faster and full attention 72.4× faster.
Key points
Speed: sparse attention up to 22.7× faster
Speed: full attention 72.4× faster
Sources
Zyphra Hierarchical Attention launches with 35× speedup over dense attention
Zyphra Hierarchical Attention (ZHA) is introduced as a path to true subquadratic attention. It claims projected speedups of 35× over dense attention and 5× over MSA at 10M tokens, rising to ~400× and ~55× at 1B tokens.
Key points
Speedup: 35× over dense attention at 10M tokens.
Larger scale: ~400× speedup at 1B tokens.
Sources
LeVJEPA 1.5 Large launches with 5.6x less compute
LeVJEPA 1.5 Large is introduced as the first step in scaling the method, closing the gap to V-JEPA on image and video benchmarks while using 5.6× less compute and over an order of magnitude less data.
Key points
Closes the gap to V-JEPA on image benchmarks.
Closes the gap to V-JEPA on video benchmarks.
Trained with 5.6× less compute and over an order of magnitude less data.
Sources
LeVJEPA 1.5 Large uses 5.6× less compute than V-JEPA
LeVJEPA 1.5 Large uses 5.6× less compute than V-JEPA and over an order of magnitude less data while closing the performance gap on image and video benchmarks. It is presented as the first step in scaling the method. Checkpoint is available.
Key points
First step in scaling the method
Trained with 5.6× less compute than V-JEPA
Uses over an order of magnitude less data
Sources
OpenAI sprint with Cathedral Systems, $172k bounty
OpenAI announced an internal security sprint partnered with Cathedral Systems and raised bug bounty rewards, offering up to $172,000 for exceptional findings. The collaboration will use OpenAI models to test defenses and expand access for trusted defenders.
Key points
Partnership: Cathedral Systems researchers will use OpenAI models alongside OpenAI teams to challenge defenses.
Rewards: Bug bounty rewards for critical findings are more than quadrupled, with exceptional findings eligible for up to $172,000, including bonuses.
Sources
Qwen3.5-0.8B weight edit improves math 6x
Qwen3.5-0.8B weights edited by GPT-6 Astra changed 0.016% of parameters, boosting math performance up to 6× and instruction following up to 3× without gradients. The approach is called Weight-Space Programming.
Key points
Edited only 0.016% of the model’s weights.
Improved math reasoning up to 6× and instruction following up to 3×.
Achieved changes without using gradients or additional training.
Sources
GPT-6 Astra edits Qwen3.5-0.8B weights, 0.016% change improves math 6x
GPT-6 Astra edited Qwen3.5-0.8B weights, changing only 0.016% of parameters, fixing reasoning loops and boosting math performance up to 6× and instruction following up to 3×, without gradients or training.
Key points
Weight change: 0.016% of Qwen3.5-0.8B weights altered.
Performance gains: math reasoning up to 6× improvement, instruction following up to 3× improvement.
Sources
TMax v1.1 released with up to 6-point TB2 improvement
TMax v1.1 is announced with model improvements of up to 6 points on TB2. The release includes public data, weights, infrastructure, and training scripts. It claims dominance on the Pareto curve for open terminal agents under 30B.
Key points
Model improvement: up to 6 points on TB2.
Public release: updated data, weights, infra, and training scripts.
Claim: dominates the Pareto curve for open terminal agents under 30B.