true or false
2.9K posts
Rust port of Small Harness is complete, and working.
I was going to hold off on releasing but I dunno, may as well just release it, Open Source, and let people play.
Going to run some security scans and then share it with the world, silly to not just build and share as I go with something like this.
If you're looking for a small harness, to use with local llms, that will get better over time but might kinda suck today, I've got you covered 😅
English
Time to wire up the cluster again 🫡
M4 Mac Mini 16 gb <-> M4 MBP 24 gb <-> M4 MBP 16 gb
EXO Labs@exolabs
EXO v1.0.70 is out. This release ships with multimodality and major enhancements for memory usage in long context use cases (e.g. @openclaw and @opencode), as well as updated model support and QOL features.
HT
hermes-lcm v0.3.0 is out — biggest release yet!🚀
What's new:
- Smart search with sort modes (recency / relevance / hybrid) + full CJK & emoji support
- Adaptive compaction that scales with backlog pressure and auto-retries on model limits
- SQLite hardening: FTS auto-repair, versioned migrations, extended lock tolerance
- Overflow recovery when assembly cap is hit
- 130 tests, ~3,800 lines of code, zero dependencies
Huge thanks to @Voscko — now a write-access contributor — who shipped the bulk of this release! 🙏
→ github.com/stephenschoett…
@NousResearch @Teknium
English
兄弟们，技术架构图终于有救了！ 🤯
以前画 Multi-Agent、RAG、Tool Call 这些架构，diagrams. net 里永远对不齐、颜色土、导出还模糊……画一次头秃一次
烟花直接放大招！开源了个超级神器 fireworks-tech-graph —— 专为 Claude Code 打造的技术图生成 Skill！
一句话就能出图，效果直接起飞：
✅ 自动识别图类型
✅ 智能语义形状 + 颜色编码（流程蓝、控制橙、数据绿…）
✅ 支持玻璃态、Neon 等高级风格
✅ 高清 SVG + PNG 一键导出
8 种图类型 + 5 种视觉风格，AI/Agent 常见 Pattern 全覆盖！
从此再也不用为画图焦虑了，太丝滑了
生产力直接拉满！
快去 star 支持烟花老师～
开源链接我放评论区👇
烟花老师@teach_fireworks
写技术文章最烦的事之一：画图。 脑子里很清楚的架构，落到 draw.io 上总是对不齐、颜色难看、导出模糊。 所以我做了 fireworks-tech-graph，一个专门生成技术图的 Claude Code Skill。 o… 用法很简单—— t 「画一张 Multi-Agent 协作图：Orchestrator 调度 3 个 SubAgent，分别负责搜索、计算和代码执行，最后汇聚到 Aggregator 输出结果，玻璃态风格」 然后它会： ① 识别图类型 → Agent 架构图 ② 分配语义形状 → Orchestrator 用六边形，Agent 用六边形，存储用圆柱体 ③ 用语义颜色编码箭头 → 蓝色主流程、橙色控制流、绿色读写 ④ 自动导出 SVG + 1920px PNG 整个过程不需要写 DSL，不需要打开任何工具，一句话描述，图就出来了。 o… 目前支持 8 种图类型、5 种视觉风格，AI/Agent 领域的常见 Pattern 全部内置（RAG、Mem0、Agentic Search、Multi-Agent、Tool Call 等）。 开源，欢迎 star 和 fork 👇 github.com/yizhiyanhua-ai…
中文
qwen3.5-35b(moe) 모델을 동적 양자화 하여, 모델 웨이트를 11.2gb로 줄여서 16gb m4 mini에서 60tps 로 컨텍스트 윈도우 245K 확보하는 데 성공했습니다.
마지막 테스트만 문제 없으면 바로 HF에 게시할 예정입니다.
Alis volat propriis@Alisvolatprop12
성공!!!!!!
한국어
gemma4(26B, MoE)에 이어 qwen3.5(35B, MoE)를 16gb unified ram에 올려서 실용적으로 쓰기위해 동적 양자화를 시도하는 데, gemma4보다 qwen3.5가 아키텍처 상 훨씬 효율적이네요.
역시 궁하면 방법을 찾아낸다는 건지...
한국어
Optimized Qwen3.5-35B on my NVIDIA DGX Spark with vLLM: 50.9 tok/s text generation with full 256K context, beating the spark-arena benchmark by 3%.
Key wins: FlashInfer attention, prefix caching, Mamba SSM cache in float16, and tuned batching. Scales to 195 tok/s at 10 concurrent requests.
Full docker-compose config: gist.github.com/wshobson/45c91…
English
@xlab_os @CherryJimbo I added cost command to get an calculation based on reported usage (don't have websockets in my demo app but should be included in the stats):
English
Ship Clouflare Containers to their edge to run anything, as easy as fly.io or Heroku deployments.
No Workers or Durable Objects experience needed.
English
M3 ULTRA 512gb 통합메모리 클러스터(x2)
DeepSeek R1 Q4: ~404 GB
Llama 3.1 405B Q4: ~203 GB
Qwen 2.5 72B BF16: ~144 GB
합계: ~751 GB
여유: ~273 GB (KV 캐시용)
이렇게 가능할 것으로 예상
한국어
모든 연구자의 꿈이었던 논문용 diagram 그려주는 모델 PaperBanana!
Dawei Zhu@dwzhu128
[1/n] Super excited to introduce PaperBanana 🍌! (PKU x Google Cloud AI) As AI researchers, we often spend way too much time crafting diagrams and plots instead of focusing on the ideas 🤯. To rescue us from this burden, we built an Agentic Framework to auto-generate NeurIPS-quality paper illustrations! 📄 Paper: huggingface.co/papers/2601.23… 🌐 Page: dwzhu-pku.github.io/PaperBanana/ Key Features: 🌟 Human-like Workflow: Retrieve 🔍 -> Plan 📝 -> Style 🎨 -> Render 🖼️ -> Critique 🔄. This ensures both academic fidelity and aesthetics. 🌟 Versatile: Supports both illustrative diagrams and statistical plots. 🌟 Polishing: Also effective for polishing existing human-drawn diagrams. Here are some example diagrams and plots generated by our PaperBanana:
한국어
true or false retweetledi
팔아서 더 오르는 건, 상상 속의 손해
사서 떨어지는 건, 실제 계좌 속의 손해
참고 - 워렌 버핏이 말하는 투자 원칙 2 가지
제 1 원칙 : 절대로 돈을 잃지마라.
제 2 원칙 : 제 1원칙을 절대 잊지마라.
Alis volat propriis@Alisvolatprop12
살까말까? 쉼호흡을 하고 30분을 참아본다. 팔까말까? 생각하지 말고 바로 실행한다.
한국어
true or false retweetledi
결국 자율주행 관련 주식은
“자동차 섹터”라는 한 칸에 담기지 않는다.
반도체·센서·AI·통신·플랫폼·보험이
하나의 흐름으로 엮인 시스템 베팅이다.
한국어