Featured
NLP
New 2026
DeepSeek OpenAI-Compatible API Wrapper
New 2026
Local push-to-talk Whisper.cpp dictation for DeepSeek Harness Web
New 2026
TUPOI: Symplectic Post-Transformer Language Model with O(1) Memory (No KV-Cache). Built by 15 y.o. independent researcher.
New 2026
🎯 AI-Powered Universal ATS Resume Tailoring Engine. Automatically reframes your experience, projects, and skills to match any Job Description into 1-page LaTeX & PDF.
New 2026
Reasoning-prefix transfer: prefill Qwen3-8B's thinking with the first 1% of DeepSeek's reasoning and measure if its answer drifts toward the source via bge-m3 cosine. A local replication of the Stolen Thoughts demo (arXiv:2608.09867). 推理前缀迁移：把 DeepSeek 推理的前 1% 预填进 Qwen3-8B 的 thinking，用 bge-m3 余弦看答案是否向源靠拢。本地可跑，复刻 Stolen Thoughts 演示（arXiv:2608.09867）
New 2026
Semantic product search API using SentenceTransformers + Qdrant vector DB to find relevant products beyond exact keyword matches.
New 2026
A powerful model compression framework for LLMs and LVLMs, adapted for NVIDIA GPUs and Huawei Ascend NPUs.
New 2026
Forensic detection, benchmarking, and explainable analysis of LLM text watermarks, provenance signals, and hidden-payload channels — with a live web UI
New 2026
llama.cpp recipe for Qwen3.8-27B GGUF on Turing RTX 6000 (24GB). Measured tok/s. Not Ada, not Blackwell.
New 2026
Notes and tools from rebuilding a cheap Proffieboard saber - hardware traps worth knowing, and a way to find out what your sound fonts actually say.
New 2026
Claude explains its work in words only it knows. An output style and Stop hook that make it write for someone who wasn't in the room.
New 2026
One llama.cpp flag unlocks +33-39% decode speed for Qwen3.8-27B on consumer GPUs. The MTP head already ships inside your GGUF. Recipe, paired benchmarks, probe tool.
New 2026
Claude Watermark Removal: Theoretical Until Proven. Nobody has cracked it. Paraphrase is not a detector.
New 2026
A curated and verified catalog of datasets for Turkish across language, speech, vision, and multimodal AI.
New 2026
Unofficial companion notebooks for language model builder app
New 2026
Runtime rank-1 refusal projection for DeepSeek-V4-Flash-0731: 757KB of directions instead of a 1.54GB weight overlay, lambda as a hot-swappable dial. Full A/B measurements on 2x DGX Spark.
Other topics
Browse other topics on Trendshift
A
N