Latest

Decision-specific micro-models are shifting routing and gating workloads away from large general LLMs, cutting both latency and cost in agent pipelines. Developers building multi-step agent workflows or RAG evaluators can drop LemonDecider in as a fast classifier, and its open training data makes custom fine-tuning feasible. Compared with general models like Qwen3.5-4B or Jev 1.14, it trades generality for a measurable benchmark lead and millisecond-level inference.

AIGCLINK@aigclinkSource

Debugging RL training is where agent evaluation meets real engineering pain, and a ~40% ceiling shows frontier models are far from automating it. Engineering teams running post-training pipelines can use the open code and configs to benchmark which agents catch data-level faults before wasting GPU hours. Compared with synthetic bug suites, its use of authentic failed runs makes the scores harder to game.

OpenBMB@OpenBMBSource

The MIT licence with commercial rights makes LongCat-2.5 one of the most permissive trillion-parameter releases, directly comparable to DeepSeek and Qwen open weights for agent and coding workloads. The 4-bit expert quantisation and chip co-design lower hardware barriers for teams running on non-NVIDIA accelerators common across China and Southeast Asia. Developers building long-context multimodal agents should evaluate it against closed APIs, with weights on Hugging Face and an API platform already live.

The heavier private-dataset weighting and confidence-aware scoring make it harder for models to game public benchmarks, shifting how leaderboard positions are earned. Teams selecting a decision model should weigh the 60x cost spread — Celeris-1.1 Decision sits on the Pareto frontier at US$14 per 1M tasks versus Drex Max v1.1 at US$102 — since token price alone understates real task costs.

Official accountArtificial Analysis@ArtificialAnlysSource

The hybrid low-resolution-plus-detail-refocus design matters because it makes 4K matting feasible on consumer hardware rather than data-centre GPUs, with a 7.4x speedup and one-thirteenth the VRAM. Developers in video editing, virtual production and VFX can watch for the promised open-source release and benchmark it against full-resolution matting pipelines. The work comes out of NTU's S-Lab in Singapore, a notable ASEAN research contribution to video generation tooling.

Gorden Sun@Gorden_SunSource

The interesting part is the method: AI agents drove the search across routing, scheduling and kernel-level changes, with a strict benchmark-before-merge loop — a reproducible pattern for anyone running inference infrastructure. Teams serving MoE models can apply the same TTFT- and TPS-targeted workflow. StepFun is a Chinese lab, so its models being served at scale on a US stack also signals deepening cross-border infrastructure ties.

Official accountTogether AI@togethercomputeSource