A Quarter the Size Still Hits 80% on SWE-bench
Inkling-Small uses 12B active parameters to approach its 975B sibling on coding and reasoning, but its factuality gap still matters.
NESTFRONTIER / ARCHIVE
All entries
Technical analysis, model releases, research, and deployment notes.
PAGE_06
291 PUBLISHED ENTRIESThirty Seconds Finally Makes AI Video Editing Less Fragile
Seedance 2.5 doubles native clip length and expands multimodal references, but its real test is whether continuity survives difficult scenes.
Open weights now come with a $20 million question
Kimi K3 is a 2.8T open-weight model with frontier scores, a 1.56 TB checkpoint, and a license that changes the commercial math.
Long videos still drown VLMs in irrelevant frames
ReToken adds one learned token to visual-language models and improves long-context image and video retrieval without retraining the whole backbone.
Production oncall is still a human job at 10% accuracy
ORCA-bench puts coding agents in a noisy six-day microservice testbed. The best hard-task RCA accuracy was just 10%.
82.7 Makes Cheap AI Coding Hard to Ignore
DeepSeek V4 Flash 0731 posts an 82.7 on Terminal Bench 2.1 and brings cheap, Codex-ready agent coding to public beta.
32 Hz Robotics Just Lost Its Billion Parameter Habit
TurboVLA removes the LLM from the robot control loop, reaching 32 Hz and 97.7% LIBERO success with 0.2B parameters and 0.9 GB VRAM.
Agent Tests Found 93.9% Attack Success
Vera tested four production agents with executable safety cases and found a 93.9% average attack success rate under multi-channel attacks.
Long context just stopped being a memory problem
HiLS-Attention learns which distant chunks deserve attention, reaching million-token extrapolation without dense attention costs.
Small Models Stop Digging Their Own Reasoning Holes
Relay-OPD lets a teacher briefly rescue a small model when its reasoning goes off course, improving math accuracy while cutting training trajectories by more than half.