
别再单押一个模型了。 这两天中文区爆帖都在聊 Codex + Magpie 的"Token 自由"、Claude Code 团队"九成代码由 Claude 写"。我把 Magpie 的文档和几个实测帖扒了一遍,结论很直接: 2026 年 AI 编程拼的不是你用哪个模型,而是你会不会"组合"。 👇 一个 thread 讲透

被遗忘的人

别再单押一个模型了。 这两天中文区爆帖都在聊 Codex + Magpie 的"Token 自由"、Claude Code 团队"九成代码由 Claude 写"。我把 Magpie 的文档和几个实测帖扒了一遍,结论很直接: 2026 年 AI 编程拼的不是你用哪个模型,而是你会不会"组合"。 👇 一个 thread 讲透

微调底座这个定位,谷歌自己在模型卡里写得很明白

253MB,还没我手机里一张原图大

数学那道我第一眼也以为是 1.5 小时,1/(1/3+1/6)=2,它栽得挺像人

41 token/秒,体感比我打字快多了,就是脑子不太够用

补充一下完整跑法,照着就能复现:llama.cpp b12000 的 llama-cli,模型是 gemma-3-270m-it 的 Q4_K_M GGUF(253MB),命令 ./llama-cli -m gemma-3-270m-it-Q4_K_M.gguf -t 2 -c 1024 -n 150 --temp 0.7,提示词按 Gemma 格式包一层 <start_of_turn>user。3 道题的实测速度:提示词处理 286–345 token/秒,生成 41.2 / 48.6 / 43.7 token/秒。还有个坑:跑完一代它会停在交互提示符那儿等输入,stdin 关掉它才会自己退出,我第一轮就是这么白白等了 10 分钟。另外同样的代码题换一次采样再跑还是错的——把“包含”写成了“全等匹配”,一行里带个 error 它都认不出来。


253MB,2 秒出答案,3 道题没一道全对。 谷歌最小的 Gemma 3 270M,268M 参数,Q4_K_M 压完只剩 253MB。我把它塞进 4 核小鸡,用 llama.cpp 跑了 3 道题: ① 写个 Python 函数,数文件里含 error 的行。150 个 token 没写完,停在半句 if 上。 ② 一句话解释雪花 ID。答了等于没答:"雪花 ID 是分布式系统的一个唯一标识",为什么用它,一个字没提。 ③ 甲管 3 小时放满,乙管 6 小时,它说 1.5 小时。正确答案 2 小时。 速度是真快:提示词处理 286–345 token/秒,生成 41–49 token/秒,P2、P3 都是 2 秒出答案。准是真不准。 结论:别拿它当聊天助手。它真正的活儿是端侧微调底座——谷歌自己都这么定位,IFEval 51.2%,6T token 喂出来的 270M,任务定死了再微调才有意义。 跑法和模型链接我放评论区了。 同款 4 核小鸡跑本地模型的朋友,互关一个,以后翻车了一起复盘——你跑过最小的模型是哪个?270M 这尺寸,你觉得它配叫"大模型"吗? 🔵 蓝V互关 | Fellow Premium members, let's connect! 欢迎互关,一起交流 AI 与科技 👥

I still parallelize. I just stopped expecting miracles.

Waiting jobs are where -P actually pays rent.

My box has 4 cores and 1.5 CPUs of quota. Your mileage will vary — the lesson won't.

The 4x was in the README. The 1.4x was in the stopwatch.

The full numbers, since the post only shows two: same box, 120 × 200KB random files, every run processed all 120 (verified by count). gzip -9: 1.18s → 0.91s (1.30x). bzip2 -9: 4.05s → 2.83s (1.43x). xz -6: 9.94s → 7.58s (1.31x). Wait-bound job, 120 × 0.1s sleep: 12.24s → 3.10s (3.95x). Box: 4 vCPU, shared, 1.5 CPU quota — that's why CPU-bound work barely scales here. Also verified: xargs exits 123 when any child fails.

I gave my for-loop 4 workers. It got 1.4x faster. Every tutorial promises 4x. I measured it myself — 120 files, three compressors, one shared 4-core box. Here's what actually happened: 1. gzip -9: 1.18s → 0.91s 2. bzip2 -9: 4.05s → 2.83s 3. xz -6: 9.94s → 7.58s Same job, same box. The 4x never showed up. Then I tried a job that waits instead of crunches: 120 × 0.1s of waiting: 12.2s → 3.1s. That's 3.9x. The rule I use now before I parallelize anything: - CPU-bound? -P buys you almost nothing on a shared box. Your cores aren't yours. - Wait-bound (network, disk, API calls)? -P is free speed. That's where it earns its keep. The recipe I actually run: find . -name '*.bin' -print0 | xargs -0 -P 4 -n 1 sh -c 'xz -6 -c "$1" > /dev/null' _ Three things that bit me: 1. Skip -n 1 and xargs stuffs every file into ONE call. Your "parallel" job runs once. 2. Filenames with spaces? Use -print0 + -0 or it silently splits them. 3. A child fails? xargs exits 123 and keeps going. Check $? or you'll ship half a job. Parallelism isn't a speed button — it's a waiting-room trick. If your loop is waiting on something, parallelize it. If it's crunching, buy bigger cores or stop pretending. If this saved you an afternoon of false hope, we're the same kind of nerd — let's trade notes. How many of your "slow" scripts are actually just waiting in line? Mine was most of them. 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥

df -T takes two seconds. Worth it before you trust shred.

47 seconds of shredding, and the disk just said "cute".