LongCat-2.5 is released under the MIT License with a 1.6 trillion-parameter multimodal MoE and about 48 billion activated per token. It adds INT4 inference on NPUs with quantized KV cache, achieving nearly double decode throughput over W8A8KV16.
Key points
1.6T total parameters, ~48B activated per token
INT4 inference (W4A4) nearly doubles decode throughput versus W8A8KV16 in stress test
Sources
- Original post
- modelscope.ai - Linked in the post
ReproBridge workflow lifts agent success to 92.6%
ReproBridge workflow lifts agent reproduction success to 92.6% and cuts API costs by ~20%. It turns papers and raw code into refined reproducible artifacts.
Key points
Success rate: 92.6% (up from 76.4%)
API cost reduction: ~20%
Sources
Zero-Query Transferable Pretend Benign Attacks – RT-TrPB generates perturbations in <56 ms
Zero-Query Transferable Pretend Benign (TrPB) introduces RT-TrPB, an amortized online generator that creates each perturbation in under 56 ms. Experiments show RT-TrPB bypasses ROBOSAC and FLD defenses with success rates near 100%. The paper also proposes STTrust, which detects over 90% of PB-style attacks and restores perception performance to near the attack-free level.
Key points
RT-TrPB creates each perturbation in under 56 ms.
RT-TrPB achieves success rates approaching 100% against consensus-based defenses ROBOSAC and FLD.
STTrust detects over 90% of PB-style attacks.
STTrust restores perception performance to near the attack-free level.
Sources
Forecasting cumulative intraoperative hypotension burden – benchmark results released
Forecasting cumulative intraoperative hypotension burden paper reports a residual-versus-persistence loss difference of –0.644 mmHg-min at the 90th percentile. The study uses subject-disjoint splits of 70,402 training, 14,270 validation and 15,575 test windows.
Key points
Residual-versus-persistence loss at the 90th percentile: 2.256 mmHg-min vs 2.900 mmHg-min (difference –0.644; 95% interval –0.814 to –0.487).
Direct quantile boosting achieved loss 2.196 mmHg-min; with additional MAP features 2.193 mmHg-min.
Sources
STAG-Net lightweight head pose estimation achieves 3.51° error
STAG-Net introduces structure-aware attention over landmark groups for lightweight head pose estimation and reaches a lowest mean absolute error of 3.51° on BIWI under cross-dataset evaluation. The method uses only 0.012–0.043M parameters and 24.1–83.6k FLOPs, running at 322.9 fps on a Tesla T4 GPU.
Key points
Mean absolute error of 3.51° on BIWI cross-dataset evaluation (FAN variant).
Inference speed of 322.9 frames per second on a Tesla T4 GPU.
Sources
Dialt Budget launch at half price, Dialt Fast 41.1% pass rate
Dialt Budget is offered at 50% of the cost of Dialt Fast with a more limited set of voices. Dialt Fast reports a 41.1% pass rate and $0.18 per call, placing it on the pareto frontier and ahead of OpenAI GPT-Live 1 and GoogleDeepMind Gemini 3.8 Live.
Key points
Dialt Budget costs 50% of Dialt Fast and supports fewer voices.
Dialt Fast achieves a 41.1% pass rate at $0.18 per call.
Dialt Fast is positioned ahead of OpenAI GPT-Live 1 and GoogleDeepMind Gemini 3.8 Live on the leaderboard.
Sources
NVIDIA Auto Ontology launches ontology layer for agents
NVIDIA Auto Ontology adds a semantic catalog built from enterprise metadata and validates queries against certified definitions and lineage. It serves the ontology to agents via MCP, CLI, and API.
Key points
Feature: Builds a certified semantic ontology from your metadata
Feature: Validates generated queries against certified definitions and lineage
Sources
Gemini 4 Carbon live on Vertex AI and Gemini Web App
Gemini 4 Carbon is now rolling out on Google Cloud Vertex AI and added to the live Gemini Web App configuration. It is mapped to the internal chrome_chai_upsell tier alongside gemini-4-argon (chrome_sencha_upsell).
Key points
Launch: Gemini 4 Carbon live on Vertex AI (Google Cloud)
Configuration: gemini-4-carbon added to the Gemini Web App and mapped to chrome_chai_upsell tier with gemini-4-argon (chrome_sencha_upsell)
Sources
- Original post
- gemini.google.com - Linked in the post
