There's another open source coding agent called Kimchi.

2026-07-15 03:42

There's another open source coding agent called Kimchi. It's built on Pi, has a /ferment mode that auto-routes tasks to different models (very similar to how I run Pii). Coding → MiniMax M3 (cheap, solid) Planning → Kimi K2.7 (strong reasoning) Codebase search → Deepseek-V4-flash (fast, dirt cheap) Multiple subagents spin up and rotate. Kimchi says 40-60% cheaper than running a single model. Free tier has a weekly quota, and like Pi you can plug in your own API keys. I tried /ferment: one sentence to build a todo app. It split the task, executed, self-reviewed, opened a PR. The output was genuinely solid.

Two weeks with SuperGrok and ChatGPT Plus

. Weekly usage is about the same. Grok Build pulls from SuperGrok's ~$40/wk credit pool. grok-4.5, roughly 50 rounds. Codex Plus GPT-5.6 Sol: 50–150 msgs/wk on paper. Grok Build has no 5-hour cooldown. Feeling it? Burn the entire weekly quota in one go. Not in the zone? Save it for later, zero pressure. Codex Plus locks you into 5-hour windows. A heavy task eats 5–11% per round. Some mornings the quota's gone by 10am and you're just… waiting. SuperGrok $30 covers chat, image gen, video, voice. Grok Build is basically a free throw-in. ChatGPT Plus $20 gets you Codex plus ChatGPT and image gen. Both available at $10 on discount plans. Which one are you subscribing to?

2026-07-15

Don't post Grok content on Meta platforms

Posted Grok stuff for a few days. Threads emailed me today asking for a selfie. Uploaded it. Permanently banned. Threads algo was actually solid. Hit 2M impressions for a few months straight. Then the traffic went up and moderation got tighter and tighter. So many words blocked now, you gotta censor with ** just to type normally.

2026-07-15

Meituan shipped LongCat-2.5

. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.

2026-11-02

SenseTime open-sourced SenseNova Loop-9B-Preview.

It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.

2026-11-02