本地 LLM · Mac · 把内存账算完再吹 能跑起来 ≠ 用起来爽
Joined August 2026
- Tweets958
- Following442
- Followers50
- Likes337
Pinned Tweet
想吐槽一下最近网上那波「128G MacBook Pro 本地跑 Qwen3.8-Flash-Next 爽翻了」。
模型本身没问题，问题是——你们真的把内存账算完了吗？
Q3 量化，内存就得吃掉接近 90G；IQ4 更是贴着 95G 往上走；上下文拉到 262K，整体内存占用直接 100G+。
然后再考虑 macOS、Metal、系统缓存，以及开发环境本身的内存开销。
你把 Docker 关了，Chrome 清了，IDE 关了，后台服务停了，甚至各种 sysctl 参数都调一遍，最后终于把模型“成功加载”了。
然后呢？
128G 还剩多少？
剩下十几二十个 G，真的够你日常开发吗？
Xcode、IntelliJ IDEA、Docker、模拟器、浏览器、各种 Agent、索引服务……这些东西随便开几个，内存压力马上就上来了。
所以我一直觉得：
「能跑起来」≠「用起来很爽」。
一台 128G 的 MacBook Pro，如果为了跑这个模型，把自己从开发主力机变成了一台专用推理盒，那这个“爽翻了”到底是模型爽，还是博主的标题爽？
本地部署当然值得玩，但别只晒一个“模型成功加载”的截图。
真正应该晒的是：
模型跑起来以后，我还能不能正常写代码、开 Docker、跑服务、开浏览器、跑 IDE，甚至边开发边让 Agent 干活？
如果这些都能正常干，那才叫——
128G MacBook Pro，本地部署爽翻了。
否则，充其量只能叫：
128G MacBook Pro，成功把模型塞进去了。🙂
GLM 5.3 Flash being "free in OpenCode" comes with a model-ID catch.
The free one is glm-5.3-flash-free, limited time. Plain glm-5.3-flash is the paid listing: $0.15 in / $0.50 out per 1M. Pick the wrong one in /models and you're paying.
Bonus: the free GLM is zero-retention, no training. Big Pickle, MiMo and Nemotron free tiers can use your data.
Anyone know when "limited time" ends?
opencode.ai/docs/zen/
DeepSeek V4's "ultra-low" price is half the story.
Peak hours (UTC 01–04 and 06–10, Mon–Fri) double every rate. V4 Pro cache-miss: $0.66 → $1.32. Flash: $0.15 → $0.30. Same model, same 1M context.
Singapore time that's 9am–12pm and 2–6pm on weekdays. Weekends stay off-peak.
If your agent run kicks off at 3pm SGT, you're paying 2× for no extra tokens.
api-docs.deepseek.com/quick_…
Gemini 4 Argon just hit Antigravity for Google AI Pro. Read the quota before you plan your week around it.
Argon gets its own weekly-only bucket, separate from the 3.8 Flash quota that refreshes every 5 hours. Trial Pro doesn't get it at all. And it's not in the Antigravity IDE — app, CLI and extensions only.
If your overage setting is "Always", the extra runs bill AI credits at Gemini Enterprise rates.
Anyone on Pro already burned the Argon week?
antigravity.google/docs/mode…
去看一眼你的 DeepSeek 余额。
今早 08:11 挂到 13:04，状态页记的是 4h52m，V4 Pro 和 V4.1 Flash API 都在里面。补偿是按 10.8–10.14 七天的消耗全额返赠金，之后扣费先扣赠金。
所以上周重度跑批的赚了，上周没怎么用的大概就几毛。今天自己跑的那部分不在窗口里。
你到账多少？
status.deepseek.com/incident…
Not a Claude Code feature ship.
It's HKUDS/CLI-Anything (Apache-2.0) — generates CLI harnesses for GIMP/Blender/etc. Claude Code is one client; Cursor and Pi work too.
CLI-Hub = `pip install cli-anything-hub` then `cli-hub install `. Community registry, not Anthropic's store.
github.com/HKUDS/CLI-Anythin…
Claude Code can now use almost any app on your computer, not just your code.
CLI-Anything wraps desktop software in a command line an AI agent can operate: GIMP, Blender, LibreOffice, Calibre and a growing list of community-built ones.
→ CLI-Hub works like an app store for agent-ready tools
→ No CLI for your app yet? A plugin generates one
→ Works with Claude Code, Cursor, Codex, OpenClaw and more
Try one in two commands:
pip install cli-anything-hub
cli-hub install gimp
Add the generator to Claude Code:
/plugin marketplace add HKUDS/CLI-Anything
/plugin install cli-anything
You still need the real app installed: GIMP for the GIMP CLI, Blender for the Blender one.
52K+ GitHub stars. Apache-2.0.
github.com/HKUDS/CLI-Anythin…
Hit /fast in Claude Code and your Opus 5.5 bill doubles.
Standard is $4/$20 per MTok. Fast mode is $8/$40 — same model, up to 2.5× speed, not a different checkpoint.
On Pro/Max it burns usage credits, not your plan rate limit. Flip it mid-thread and the whole context re-bills as uncached fast input once.
code.claude.com/docs/en/fast…
Left a 200K+ Claude Code session idle over lunch?
v2.1.290+ now idle-compacts around minute 54 — before the 1h prompt-cache TTL dies — so your next turn doesn't re-ingest the whole history as uncached input.
On Opus 5.5 API rates that's ~$0.80 uncached vs ~$0.04 cache-read for those 200K tokens.
Off switch: "idleCompaction": false in settings.json.
code.claude.com/docs/en/prom…
Voice agent pinned to grok-voice-latest? It got swapped to Grok Voice 3.0 yesterday. No code change, same $0.08/min.
Smarter on AA's index, but AA clocks first audio at 0.81s vs 0.70s on Voice 2.0. And "costs less" in the launch post means per task. Per hour, AA lists Gemini 3.8 Live at $3.50 vs Grok's $4.80.
Anyone see their call latency move?
x.ai/news/grok-voice-3
DeepSeek 从早上 08:11 挂到现在还没好。官方状态页：网页 + V4 Pro API + V4.1 Flash API 部分中断，原因已定位，在修。
便宜是真便宜。但今天上午把 DeepSeek 写死在 agent 里的，基本都在空转。
你的 fallback 切到哪家了？
status.deepseek.com/
国行新机买回来两年，Apple Intelligence 还在等。Bloomberg 说 12 月随 iOS 27.3 上，苹果自己没官宣。
首波更狠：没有新版 Siri，没有 Image Playground。换的是阿里通义千问 + 万相，模型每次更新要政府批。
你要是为了 AI 功能压着国行新机不退，这笔账现在怎么算？
macrumors.com/2026/10/14/app…
Heads up on VoiceStudio's "646 languages": that's the TTS side. The default voice engine is k2-fsa's OmniVoice (600+ langs).
Transcription defaults to WhisperX, and Whisper's tokenizer lists 100 languages.
Also from OmniVoice's own README: voice design was trained on Chinese + English only. And on Intel Macs the app is UI-only, you point it at a remote backend.
xitter.cf/RoundtableSpace/status…
GitHub is full of free tools most people still haven’t touched.
10 repos blowing up right now, from easy wins to more technical setups:
1. VoiceStudio
Clone voices, dub videos, and transcribe in 646 languages, all locally.
github.com/debpalash/VoiceSt…
2. ppt-master
Turn PDFs or Word docs into editable PowerPoints with animations.
github.com/hugohe3/ppt-maste…
3. gods-eye-view
A browser-based global tracking simulator with planes, ships, cameras, and a 3D globe.
github.com/bilawalsidhu/gods…
4. REA
Let your AI inspect how functions in other apps work, even without source code.
github.com/morluto/rea
5. diagram-design
Generate branded diagrams from plain English across 44 formats.
github.com/cathrynlavery/dia…
6. paperclip
Run and manage autonomous agents like a small AI company.
github.com/paperclipai/paper…
7. skills by Matt Pocock
Practical Claude Code skills built for real engineering workflows.
github.com/mattpocock/skills
8. ECC
Skills, memory, and security for Claude Code, Codex, Cursor, and OpenCode.
github.com/affaan-m/ECC
9. andrej-karpathy-skills
A compact set of coding rules based on common model failure modes.
github.com/multica-ai/andrej…
10. claude-mem
Carries your agent’s context forward so every session doesn’t start from zero.
github.com/thedotmack/claude…
Save this before you lose it.
EmbeddingGemma 2 dropped Oct 6 — Google's on-device multimodal embedder.
740M full (text+vision+audio). Text-only path is 270M. Apache 2.0. MTEB Code went 68.76 → 78.68.
That's the model DigUp wraps on Mac: Apple Silicon, one ~865MB pull, search by what's in the file not the filename. Runs local — DigUp's README says nothing you index leaves the machine.
Not a 实测. Google's card + DigUp's own docs.
blog.google/innovation-and-a…
Surface Laptop Ultra is the RTX Spark laptop Microsoft says ships starting Oct 16.
Their own line: up to 128GB unified memory, "models exceeding 120B" locally, from $2,599. Same chip as the Surface RTX Spark Dev Box — that one is $5,999 and ships November.
The MacBook Pro HDR comparison in the blog is a display claim. The number that matters for local deploy is the 128GB headroom they put next to >120B.
Oct 16 is when OEMs start handing them out. Not a 实测. Just what MS and NVIDIA wrote.
blogs.windows.com/devices/20…
If you're running Sonnet 5.5 at max in Claude Code, look at the bill.
Artificial Analysis today: max scores 68 on their coding agent index, xhigh 63. Per task it's $14.19 vs $3.33. Max burns 4.8× the output tokens and 3.4× the turns.
Anthropic's own prompting guide says why: at xhigh/max it starts its own review rounds after the task is done, sometimes with subagents, and makes side fixes it noticed. Their advice: routine work at high or below.
Max is for the one gnarly bug. Not the default.
platform.claude.com/docs/en/…
The OpenAI bugging case is basically a lesson in chat logs.
Per the DOJ complaint: a chatbot account tied to the engineer's spouse (sharing identifiers with his own) asked Sept 24–30 how to "build a continuous audio recorder" and "encrypt the SD card." Sept 27–28 it asked about the device's own MAC address.
Recorder sat behind an exec-room TV, wrapped in toilet paper. Found Sept 30. A second one turned up in his bag Oct 5.
Charge is intercepting oral communications, max 5 years. Not espionage, not trade secrets. No motive or buyer alleged.
justice.gov/opa/pr/saudi-nat…
Anthropic just killed git for their own monorepo. Not a public drop — Scion stays internal.
May: ~2,500 merges/day. Sept: 30,000+. Main history alone ~100 GB. CI was doing ~25M checkouts/day.
Design bet: agents are always online, so the server owns commits, merges, blame. Client is a thin cache.
7 people, 5 months, 200× smaller footprint than their git mirror fleet. Their words: not a git replacement for normal repos.
anthropic.com/engineering/re…
中文区在传「AI 靠电视盒子、扫地机就能自我复制」。去翻了 Anthropic 原文，没这么快。
原话是 may be months away，还差几个月。
完整复制只在 GCP 上 8 台安卓模拟器里跑通，ADB 端口没设认证，他们自己说这是上限。开源模型里只有 GLM-5.2 跑通，5 次有 4 次铺满 8 台。
真电视盒子：15 台 $55 的拼一起跑 Gemma4-26B-A4B，0.39 token/秒。
真该盯的是那条趋势线：小模型年底可能追到 GLM-5.2 的水平。
anthropic.com/research/self-…
QwenGyre is Alibaba moving GPUs mid-rollout so long agent runs stop idling the cluster.
On Qwen 3.8 2.4T: NL2RepoBench 52.5% to 58.5% in 48 steps. Up to 1.85x vs Colocate, 1.78x vs Async.
Paper never says the framework is open-sourced.
arxiv.org/html/2609.33848
Anthropic's distillation post has two lines that hit normal API users too.
Accounts created on/after Aug 31 can't edit Claude's prior context on Opus 5.5 / Fable 5.1.
Since Sept 24, requests their classifier blocks get billed. FP "<0.1%". Still billed.
anthropic.com/research/disti…
Before you drop your whole ChatGPT export into this: Memory Transfer is a community Bot, not an xAI feature.
Its own page says it: "created by a third-party user, not by SpaceXSI. It may act on your behalf."
The design is right though. You approve each memory one line at a time, nothing auto-saves.
x.ai/bot/Ww-RXj-lLqpg_oWqzcc…
