Everybody and their mums building the infra. So who's building the damn apps?

Incredibly weird take. Just customise your harness to fit the model as much as possible?

Interesting, Opus starting from 4.7 seems to be trained on the apply_patch tool call that GPT uses.

Can it run doom?

Normally an open source tool getting acquired by a larger corporation usually doesn't bode well. But cloudflare has a good reputation amongst open source projects, so I'm cautiously optimistic!

Any terminal benchmark scores out yet for Gemma 4 12B?

Pretty pretty please open source 35B and 27B next!!! 🥹

👏👏 Introducing Qwen3.7-Plus — a multimodal agent model that unifies vision and language into one versatile agent foundation. ✅ Multimodal interactive hybrid agent: unified GUI & CLI operation across visual and text tasks ✅ Versatile coding agent & productivity assistant with full-modality input ✅ Visual Agent: perception, reasoning, grounding, and search-augmented QA ✅ Cross-harness generalization across diverse agent frameworks One model. Sees, thinks, codes, acts.🙌🙌 Now available via API on Alibaba Cloud Model Studio. Try it — let us know what you build.😎 🔗🔗⬇️⬇️ Blog：qwen.ai/blog?id=qwen3.7-plus Qwen Studio：chat.qwen.ai/?models=qwen3.7… API：modelstudio.console.alibabac…

If you work with multiple different models, you can almost certainly map them to people you've known. GPT: principled, meticulous, never wants to be wrong, but replies are slightly autistic. Opus: gets your vague brief immediately, but occasionally goes rogue and destroys half the work. Gemini: the creative one who just wants to rest and vest. Qwen/Deepseek: the intern grinding 80hr weeks who never quite hits the mark.

People of Pi! Made a Bun-native extension pack, with web access, subagents, revamped core tools, ANSI-compatible themes, fzf-style completions, Telegram mode, and more. Called it Pim (Pi IMproved) 😏. Its goal is to improve the out-of-the-box experience for both users and agents, without sacrificing composability with other Pi extensions. Pim with Qwen3.6-35B managed to average 37.8% (Peak of 41.6%!) on Terminal-Bench 2.0 over 3 full runs. Pretty damn amazing that a local model running on my MacBook can rival the performance of Claude Code Sonnet 4.5 (40.1%) and outperform Codex GPT-5-Mini (31.9%)... Open source and MIT. Try it out and tell me what you think: github.com/AaronCQL/pim-agen…! Ps, huge thanks to @badlogicgames @mitsuhiko for creating and making Pi so easy to customise, it was an absolute joy to work with!

TIL Jira MCP costs 12K tokens on startup, even when you don't use it... Audit your installed MCPs folks.

We finally solved a pet peeve of ours: simulated balance changes on Jupiter Wallet for Jito bundles show accurately across all txs now! Thanks @PierreArowana for the PR at github.com/jito-foundation/j…, and @0xTsathir for helping out.

Who approved this epileptic animation?

Ok, claude legit feels kinda lobotomised now

One of the few things I'm extremely proud of in Jupiter is RTSE. Working with the team day in day out to fine tune how we best estimate expected output and slippage, scouring over tons of data to get the most optimal calculation - the magical point that maximises output amount against minimising swap failures. This article is a great read and sums up quite succinctly the battles we've had to fight over the past few years. Special shoutout to @melvinzzy, @gn_dnomsed (and many more behind the scenes) - the brains behind it all. And special shoutout to all of our users, for always giving us the feedback that we need to improve on the system. Even now, we're still iterating and making it better. If figuring out slippage is still a chore to you, come speak to us!