Petr Glaser

@AdamRackis I took it as inspiration and while adding @tan_stack to @modern_js I am also working on proper RSC support in Tanstack way there. Looking forward for feedback from @ByteDanceOSS team for my PR and also for your blog post!

I'm not usually one to critique TanStack but you remember that RSC blog post everyone loved? Turns out it did a poor job showing just how amazing the implementation actually is; it's so much better than what they showed Working on my own post now 😈

@theo I use it, because I prefer Schaltwerk workflow over what T3 Code chose. T3 Code has everything in side panel, but Schaltwerk has tabs (like browser) for projects and sidepanel for chats. Also Schaltwerk supports "subagents" view. Would love to share my workflow on call.

Are you still using the CLI versions of your preferred agent instead of desktop apps like Codex App, Conductor, or T3 Code? Tell me why below. Genuinely curious.

@capjmk @fxckdev_ @orcdev @zeropsio Yes, you can manage everything dynamically. I do it for my Agentic Engineering Platform (ai.bleeding.dev) and Zerops team is awesome and helpful! :)

@fxckdev_ @orcdev @zeropsio @BleedingDev Okay but is it possible to tear down the containers, databases etc? What databases do you support?

Shipper Club just unlocked its first Diamond Sponsor 💎 @zeropsio 🔥 one of the cleanest platforms for deploying and running apps is now backing the builders inside the club huge thanks to the Zerops team for believing in this 💚 this is exactly what Shipper Club is about real builders. real tools. real taste. If you've been thinking about joining… it is time 🚢

@tusharmath Looking forward to that. Hope that I'll be able to build headless agent via API / SDK that can use skills. I'd love to integrate it into my platform (I already integrate Codex with others coming). :)

ForgeCode 3.0 is shaping up really well. The core is written in Rust and is nearing completion. It provides a robust API to build extensions on top of it - in any language.

@ScriptedAlchemy @iotcoi might have some tips. My bet would be GLM 5.1, so that you fine executor model, but keep ultra smart one in the cloud (GPT 5.5), so that they collab together in hybrid way, just like you already do with 5.4 Mini.

I have a B200 cluster. Looking to fine tune a model. What’s the best OSS model that’s near GPT 5.4 or opus 4.7

@jullerino @r_marked @t3dotchat I am not able to see it. Is it possible that I disabled it or something? I remember I saw it in very early version (like 0.0.3) but now nothing.

@BleedingDev @r_marked @t3dotchat it's not removed? i am rebuilding the entire orchestrator rn though to support powerful features

New model picker live in latest nightly! ⌘+Shift+M or /model to open ⌘+1..9 to quick select, or navigate with arrow keys 🔍Search ⭐ Favorite Thanks chrono for the initial work copying @r_marked's design from @t3dotchat

@elmoche_ @mudler_it @no_stp_on_snek Which GPU?

Okay. 16gb vram. Llama.cpp @no_stp_on_snek fork. @mudler_it Qwen 3.6 MoE 35B Apex i mini.gguf with 82k ctx and ctk q8 and ctv turbo3 Prompt eval: 3700 tk/s Prompt write: 114 tk/s Unsloth’s UD IQ3_S quant. (Same model and params) Prompt eval: 1200 tk/s Prompt write: 90 tk/s

@bstnxbt Would love to see local Qwen with at least 128k context, so that it is usable for coding agents. Any chance you can benchmark big context (also filled up, not just available) and big output generation?

DFlash v0.1.4 : custom Metal verify kernels for quantized Qwen3 hybrid models, plus significant peak memory reduction at long context. M5 Max 40-core GPU, 64GB, stock mlx_lm baseline: Qwen3.6-35B-A3B-4bit: ► @ 1024 · 138.3 → 300.3 tok/s (2.20x) ► @ 2048 · 135.6 → 246.4 tok/s (1.81x) ► @ 4096 · 134.5 → 208.4 tok/s (1.56x) ► @ 8192 · 133.2 → 177.4 tok/s (1.33x) Qwen3.5-27B-4bit: ► @ 1024 · 33.5 → 79.0 tok/s (2.37x) ► @ 2048 · 33.1 → 70.2 tok/s (2.12x) ► @ 4096 · 31.5 → 55.7 tok/s (1.77x) ► @ 8192 · 33.9 → 45.3 tok/s (1.34x) Working on making this usable for agentic workloads goal is to never drop below baseline at any context depth. LLM decode is memory-bandwidth bound. M5 Max runs at 614 GB/s, that's 1.5x more than M1-M4 Max (400-410 GB/s). Results will vary on lower bandwidth chips.

@thekitze Decided to go for Slidev, by far most versatile.

what's the best web based slides framework in 2026? i wanna migrate from keynote

@kcosr Don't sleep on GitHub Copilot CLI, it's also really really good!

After a few months’ hiatus, I’m back to using pi as my daily driver for interactive sessions. Around the time GPT-5.2 was released, compaction in Codex got really good, and it felt like the best harness for maintaining coherence over long implementation sessions with OpenAI models. Now that I’ve split off implementation into a headless orchestrator, I'm using pi for up-front interactive planning. I’m excited to try out some extensions and learn how to use forking and branching effectively.

@kcosr Looking forward to that! :)

Orchestration has been achieved.

@jun_song Will it be possible to convert into LiteRT-LM, so that I can run it inside Edge Gallery on my Pixel?

SuperGemma4-e4b-abliterated 프리뷰 마지막 양자화 진행중, 작업은 완료되었습니다. - 0/100 무검열 - base모델보다 훨씬 뛰어난 성능 - 스마트폰에서도 구동 가능한 사이즈 ~8gb 측정기준 : GPQA, MMLU 등등 공인벤치 항목 종합 측정 팔로우하고 조금만 기다려주세요.

@jun_song On it, about to buy server with 768 GB RAM and 288 GB VRAM. :)

opus-4.7은 성능 저하 이전의 opus-4.6과 큰 차이가 안날거라고 생각해요. 하지만 가격은 더 비싸지겠죠. VRAM을 구매하세요.

@ScriptedAlchemy 288, but if we collab, it has enough slots for 768 (8 GPU slots).

@BleedingDev How much vram total?

I’ve been fine tuning a model locally. Today i broke and got my hands on a h100 instance I can ssh into. Totally worth a dollar an hour.

@tannerlinsley @dosco @tan_stack We love it. We are using it to build skills for our design system and it's something really smart. How do you do it that you always create something better that I feel nobody else thought about? 🤔

@BleedingDev @dosco @tan_stack Yep. Go check out tanstack.com/intent. It's a "meta" registry on top of NPM (not perfect, but let's not make new problems for ourselves). Provenance, receipts, versioning, etc

agents out there are struggling with these naive and brittle implementations the paper barely touched the real key to unlock agents, RLM with GEPA optimization of the core agent + harness code with a solid eval harness.

@0xClandestine Oh, that would be really great to use this! Looking forward to it. 💙

Heterogeneous acceleration on Apple Silicon achieved. ANE + GPU running in parallel. Mirror SD with DFlash, ported to MLX — targeting ANE + GPU simultaneously. The M-series was designed for this. We just hadn't unlocked it yet.

@kky_kongjiacong Probably. I'm yet to find how these puzzles come together with Sparkling / Lynx. :)

@BleedingDev modernjs.dev/guides/advance… Modern.js cross-project BFF probably helps for the `share backend` part?

Business teams care more about AI now, and much less about frameworks. But still, I wish more people would try Modern.js.☺️ Hono + Rsbuild + React Router, with progressive RSC + CSR support and mf integration. modernjs.dev/index.html

@dosco I agree with this, just had a discussion with my friend who misses it. I think that @tannerlinsley and @tan_stack has some ideas about this with Intent!

@BleedingDev i strongly believe skills should be managed like code not put in some different store where they can get out of sync this way they are upgraded with the software