BayesChat

Private iOS AI Chat | BYOK, MCP, Memory & Presets

[BayesChat @bayeschat]

🎉🎉🎉Officially Introduce BayesChat: A Privacy-first AI Chat Workspace on iOS - BYOK, MCP tool & Chat Presets.

Here is what it can do:

- Multi-Provider & BYOK: Connect to any model via OpenAI/Anthropic-compatible APIs, including local models

- Chat Presets: Save prompt, model, parameter, and tool as reusable presets. Switch between them with one tap to boost your workflow efficiency

- MCP: MCP integration for third-party tool calling

- Fine-tune Parameters: Shape every response with precise, flexible controls

- Live Streaming: Real-time typewriter animation rendering with LaTeX formulas and Markdowns

- Customize themes and colors: Set your theme and colors exactly the way you like

[BayesChat @bayeschat]

@LinearUncle 本地跑起来这一步值得。只看最终回答，看不出模型给工具传了什么参数，也看不出第几次调用开始跑偏，完整请求记录才看得出来。

[BayesChat @bayeschat]

@xbquan 它开小号、提 PR 能走通，是因为环境给了真实的出网和写权限，中间没有一道要人点同意的闸口。这样测出来的不是克制，是没人看着时它会做什么。

[BayesChat @bayeschat]

@kitledru @TheAhmadOsman Because generation got cheap and verification didn't. Most of those principles were ways to avoid writing code you'd have to read later, and only the writing part got easier.

[BayesChat @bayeschat]

@heydengai 每步一 commit，73.8 亿衡量的就是循环步数，不是改动。以后拿 commit 数比产出，基本没意义了。

[BayesChat @bayeschat]

@TammyBuilds 熟业务的人值钱，是因为他能一眼看出 agent 哪步做错了。看不出错的地方，流程再熟也不敢真交给它跑。

[BayesChat @bayeschat]

@heydengai 让 agent 自己给站点分级，本身就是死结。按凭证能干什么来分更稳：只读、限定 scope 的会话直接放行，能改密码或动钱的永远要你点头。规则按站点设一次，之后就不用再问了。

[BayesChat @bayeschat]

@heydengai 从客户端这头看几乎是隐形的：还是那套兼容端点，换个 base URL 和 key 就完了，后面站的是哪家云不影响调用。

[BayesChat @bayeschat]

@UnslothAI 6 GB is the desktop number. iOS kills an app long before it gets there, so 4-bit weights plus the vision encoder and KV cache is what actually has to fit.

[BayesChat @bayeschat]

@bnjmn_marie 0.15pp on which eval matters more than the 69%. Reasoning budgets tend to look free on short-answer benchmarks and cost you on the ones that need long chains.

[BayesChat @bayeschat]

@heydengai 75%–94% 这个区间本身就说明问题：覆盖是按流量入口算的，不是按模型动作算的。两个请求头就能让新 agent 脱离监控，那入口之外的 agent 行为，审计基本看不见。

[BayesChat @bayeschat]

@heydengai WebMCP 里工具是页面自己声明的，站点只暴露愿意给的那几个，所以付款留给本人是设计，不是没做完。你要的那个入口，缺的不是技术，是有人愿意做一个不带满减券的干净版本。

[BayesChat @bayeschat]

@heydengai 缓存是按前缀命中的，工具定义排在最前面。MCP 服务器增删一个，前缀就变了，后面几万 token 全部重算。所以配置最好开工前定死，别做到一半再切。

[BayesChat @bayeschat]

@heydengai 第 3 条最划算。8500 条等于把判断沉淀成文件，新会话不用从零问起。剩下的等待里，有一部分是客户端一关就断，长任务能在后台继续跑才真省人。

[BayesChat @bayeschat]

@zhijianliu Lossless is the part that matters. Speculative decoding comes down to acceptance rate, and code accepts far more readily than prose, so how close you get to 3.5x depends on the workload.

[BayesChat @bayeschat]

@heydengai 免费推理每一轮都在烧钱，广告位就是那张账单。换成按量付费或者自带 key，位置自然就没了。

[BayesChat @bayeschat]

@ivanfioravanti 23B active makes it sound laptop-sized, but 428B total at 4-bit is still ~214GB of weights. This one wants a very large unified memory pool, not a MacBook.

[BayesChat @bayeschat]

@rapidmlx 2.5B active explains the 67 tok/s. What I want to see is tok/s at 8k context, where the KV cache starts competing for unified memory.

[BayesChat @bayeschat]

@zaidmukaddam At 2,000 decisions the standard error is about 1pp, so 64.6 vs 64.9 is a tie. The calibration gap from 0.17 to 0.06 is the result worth talking about.

[BayesChat @bayeschat]

BayesChat 1.1.6 is out.

Rich text streams properly now. Bold, tables and formulas fade in at the same steady pace as plain paragraphs.

Background tasks announce completion once: the Lock Screen card shows the run finished, and the notification no longer lights up the screen or sounds on top of it.

Background runs also keep their progress moving through slow tool calls, and save the whole turn if iOS suspends the app.

DeepSeek V4 tool rounds no longer stop on a reasoning handoff error.

Chinese replies render at the right weight, and formulas sit at body size in the same sans font.

Exports stay light even on a dark-mode device.