Tag
ChatGPT for SNES is an unofficial experiment by Benj Edwards that runs a tiny ~2.2M-parameter language model on the Super Nintendo's Super FX chip, offering a ChatGPT-style chat interface with an optional fact book of 150,000 Wikipedia article openings stored in MSU-1 data.
Aurora2-400M-A20M is a tiny decoder-only sparse MoE language model with ~400M total and 19M active parameters, pretrained from scratch on 100B tokens in about 36 hours on a single RTX Pro 6000 GPU. It features top-1 routing with 32 routed plus 1 shared expert per layer, grouped-query attention, hashed n-gram embeddings, value residuals, and other architectural extras, released as a base (non-instruction-tuned) model on Hugging Face.
SenseTime's SenseNova team officially releases SenseNova-Loop-9B, an open-source ~9B parameter model that reuses shared middle layers via a looped architecture with manifold-constrained hyper-connections (mHC) to gain deeper computation with fewer parameters, targeting complex reasoning, coding, tool use, and multimodal tasks, with AWQ and GGUF quantized weights now available.
A blog writeup describing how the author post-trained an autoregressive language model to autocomplete and remix 140-character JavaScript 'Dweets' from Dwitter.net, inspired by classic demoscene sizecoding effects like HellMood's 'Memories'.
bootgpt 是一个完全容纳在 510 字节 MBR 中的神经网络运行时,加载下一扇区的 11.5KB int8 量化字符级模型,在 386 的 16 位实模式下无操作系统运行,用于在开机时生成虚构的 Linux 启动日志。
DiffuSpace released Clover-35B-A3B-Preview, an open-source block-diffusion language model with a 35B MoE backbone (~3B active params), 128K context, 680 tok/s on GB300, 73.6% SWE-bench Verified, and Apache 2.0 licensing.
Metis-1.6 technical report introduces Praxis (2.94B) and Logos (10.23B), two MoE language models with a hybrid Mamba-2/attention stack built by a single researcher on ~1.02T tokens, achieving Pareto-efficient math performance while openly detailing training failures and fixes.
The paper introduces Latin Modern BERT (LAMB), a ModernBERT-based encoder pretrained on over 74B Latin tokens with morphologically informed masking, alongside two new Latin corpora (Acervus and Corpus Trinum), and shows it outperforms 15 existing encoders on 16 of 17 benchmark subtasks.
Nanbeige released Nanbeige4.3-33B, an Apache 2.0 teacher model with a 512K context window intended to train smaller agent models like an upcoming 3B variant, reporting strong Terminal-Bench 2.1 results but mixed performance on SWE-Bench Pro.
Hugging Face released Carbon-X, an open DNA language model that answers questions about gene sequences in plain English, scoring 66.7% on DNA benchmarks near GPT-6.1 Sol at roughly 1/1,800 the cost per problem. It runs on a single 80GB GPU, potentially making large-scale variant classification dramatically cheaper for hospitals and researchers, though it is not yet clinically validated.
Carbon-X is an open 14.8B decoder-only model that jointly processes raw DNA and text via a shared 6-mer token stream, achieving 66.7% average on five DNA benchmarks — nearly matching GPT-6.1 Sol — with Apache 2.0 weights that fit on a single 80GB GPU.
Recursive released Flint, a 40B total / 2.5B active mixture-of-experts coding language model on Hugging Face, scoring 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, with SGLang serving kit and quantized variants.
Mercury 3 是目前发布的最强扩散式大语言模型,宣称以 1,000+ tokens/s 的速度运行,比竞争对手快 4–8 倍,并在性能上超越 GPT-6 Luna 和 Claude Haiku 5.5,实现 40 倍扩展而不降低速度。
XtrCom 释放了 XT-Luna 1B v1,这是基于 MiniCPM5 1B 面向智能体(agentic)和编码任务微调的模型,并在模型卡中提供了与官方 MiniCPM5 1B SFT 及 Qwen 3.5 4B 的对比基准。
DeepSeek V4.1 Flash (1017) is announced as the new leader among open-weight models, topping current open-source LLM benchmarks.
Gen-Verse 发布了 DiscoveryRSI-27B,一个 27B 指令微调语言模型,用于递归式科学假设开发、评估、修订与归因,配套 DiscoveryRSI 研究智能体框架支持生命科学研究工作流。
DeepSeek V4.1 Flash 1017 已上线 OpenRouter,提供 API 定价与可用提供商信息。
Muse Glimmer 是一款 30B 参数的开源多模态智能体模型,通过 4-bit 量化可在消费级硬件上本地运行多步推理与工具调用,并在同尺寸级别 agentic 基准上超越 Gemma4-31B 和 Qwen3.6-27B,现以 Apache 2.0 协议发布。
Stanford HazyResearch 提出 Next-Concept Modeling (NCM),一种既非自回归也非扩散的粗到细并行文本生成方法,据称比强扩散基线快 10–31 倍且训练 tokens 更少;不过实验规模仅到 725M 参数,能否扩展至通用 LLM 规模仍待验证。
首尔大学团队提出KBS(Korean Benchmark Suite),包含Ko-ARC、Ko-GSM8K等六个基准共9,396个条目,采用非纯翻译方式构建韩语评测集,并评估了韩国自研模型、多语言开源模型及闭源API模型,指出评测结果高度依赖任务构建、评分规则与提示方式。