The archive · Page 7 of 7
Where architecture, training tricks, and eval methodology meet the marketing layer — separating durable progress in foundation models from leaderboard theater that quietly falls apart under load.
145–159 of 159 articles · Newest first
An arXiv paper shows the embedding learning rate accounts for most of μP's advantage over standard parameterization, and a single scaling fix recovers the bulk of the benefit.
Alibaba's dense Qwen3.6-27B outperforms its MoE sibling on coding benchmarks, trading predictable inference latency for a larger memory footprint than sparse alternatives.
Compare DeepSeek, Qwen, Kimi, Doubao, ERNIE and GLM through dated benchmarks, license terms, context windows, API pricing and practical workload fit.
A local R1 memory plan needs the chosen weight artifact, the distill's own configuration, and the intended context length.
Fish Audio's open-source S2 TTS hit SOTA in March 2026 with sub-100ms latency and 80+ languages, challenging ElevenLabs' commercial pricing and exposing licensing gaps.
Epoch projects that public human-generated training text could be fully used between 2026 and 2032, depending on scaling and overtraining.
TimesFM 3.0 adds native multivariate forecasts, but its downloadable weights are not licensed for commercial production use; production use runs through Google Cloud.
Google reports different Gemini 1.5 Pro recall results for single-needle and multiple-needle retrieval at one million tokens.
DeepSeek's V3 and R1 models match GPT-4-class performance using a fraction of the compute through architectural innovations in Mixture of Experts, attention compression, and reinforcement learning, demonstrating that training efficiency may matter more than raw hardware scale.
Claude's web search delivers real-time retrieval inside the reasoning loop with mandatory citations, domain filtering, and dynamic HTML processing that cuts token use by 24%.
Million-token context windows let you load entire codebases in one pass, but models lose coherence before their limit. Here is what benchmarks and failure modes reveal.
A hosted instance saves you some setup work, while those choices remain yours.
WiFi DensePose maps measured radio-channel data to dense human pose estimates in a research setup.
SWE-bench, LiveCodeBench, EvalPlus, and Aider measure different coding tasks and should not be merged into one ranking.
Which LLM to pick for OpenClaw in 2026: Fable 5, Opus 4.8, GLM-5.2, Kimi K2.5, and Gemini 3.1 Pro ranked by use case, benchmark evidence, and pricing.