Qwen3-32B-MLXをMacBookProで走らせる
LMStudioで実験
とりあえず飛行機の中でも試せるようにQwen3-32b-mlx(4bit)をダウンロードしておいた。
なかなか賢い。たしかにGPT-4o並というのはそうなのかも。
これがローカルで、ノートPCで動くってすごいよな



「UberEats配達員を主人公としたミステリー短編小説を書いて」と頼んでみた。以下結果
タイトル:『濡れた配達と消えたデータ』
第一章:雨の夜の注文
東京・新宿の高層アパートに、大雨が激しく打ちつけていた。配達員の神谷健太は、UberEatsアプリにポップアップした注文をチェックする。受け取り先は「新宿区西新宿1-23-5 本厚町ビル804号室」。不思議に思ったのは、注文内容が「コーヒー(ブラック)1杯」であること。価格は通常の50分の1で、配達先も閑静な住宅街に突き刺さった一軒。健太は「この雨で誰かがただコーヒーを飲もうとしているのか?」と首を傾げながら、アクセルを踏みしめた。
第二章:不協和音の配達
804号室に到着した際、ドアを開けたのは五十代くらいの男性だった。その手には小型のトランクが握られ、不自然に緊張した表情を浮かべている。注文確認の声には「ああ、はい」という不自然な吞気に続き、「どうぞ」一句でドアが閉ざされた。健太は部屋の中が見えたわずかな時間に、男性の脚元に赤黒い液体をこぼしたような跡をとらえた。
第三章:落下物と音の追跡
20分後、健太はドライブレコーダーに録画された「12時3分、近所からサイレンが聞こえる」音声を耳にした瞬間、再び本厚町ビルへ向かわねばならない。今度は「マンション1階のゴミ置き場で異様な光が漏れている」ことをユーザーH27456が報告していた。健太は服を引っ張って、配達用バッグのチェーンに何かがついていないかと探る。そこには先ほどの男性が落としたと思しき、赤黒い液体に染まったSDカードと1万円札が束ねられた封筒。
第四章:交差する配達者
健太は次の注文で、別の配達員・佐藤に声をかけた。彼女は「神楽坂で女の人が泣きながら『警察に助けを…』と叫んでいた」と話した。健太は「不完全なデータもあれば、全体が見えるかもしれない」とSDカードの写真を彼女に送った。2人はドライブレコーダーを共有し、暗号解読アプリを使ってデータの断片から「西新宿804への詐欺的な注文は、資金移動のカモフラージュだった」ことを発見する。
第五章:消えた雨と真実の配達先
健太たちはデータを匿名で警察に送る。次の夜、彼は「新宿~渋谷の往復10件分のお礼に」と、謎を解いたというカップルから「手取り20万円」のボーナスを依頼されたが断った。その後、健太はアプリから「あなたの配達履歴を確認してください」という連絡を受け、ログインすると自分が乗っていた車の動画が表示された。しかし、その先にあるのは「配達員としてのあなたは必要なくなるかもしれません」という警告だった。
Epilogue:朝焼けと次の注文
健太はタクシー会社の営業所で「雨の日よりずっと長生きできますよ」と笑いながら席を立った。その日の最初の注文は「神楽坂1丁目8-3 2F 『ありがとうのメッセージとコーヒー1杯』」。健太は心の中で、「次の配達も、どこかに真実があるのかもしれない」と呟いた。そして、再び雨が降り出してきた中を微笑みながら走り出した。
テイスト:
この小説は日常の「配達」を通して、表と裏に存在する社会の陰影を描く。ヒロインたちはそれぞれ独自の視点で、表面的な「食事」という単位に潜む真実を解き明かす。結末は完結しないが、読者に「次の注文」への好奇心を喚起する構成とし、ミステリーの完結性に挑戦しています。
もうちょっと賢いかもしれないQwen3-30B-A3B-MLXで試してみる
タイトル:夜の配達、そして鍵
深夜10時。UberEatsのドライバーである佐藤健太郎は、スマホの通知音に目を覚ます。この時間帯には通常、注文が少ない。しかし今夜は例外だった——「高田馬場駅周辺、23時までに届けてください」という依頼が入った。
配達先は「千駄ヶ谷3丁目」という住所。地図アプリで確認すると、そこは都会の奥に位置する静かな住宅街だった。健太郎は自前のバイクで向かうが、その場所には誰もいない。注文者の名前は「山田美咲」と記載されており、電話番号も提示されていたが、折り返しの連絡はなかった。
「配達完了」とアプリに報告した直後、健太郎のスマホが震えた。新たな通知——「配達先の扉に鍵を忘れました」というメッセージが届いた。
急いで現地に戻ると、建物の脇に小さな箱が置かれていた。内容を開けると、「緊急時にどうぞ」と書かれた紙が入っていた。箱の底には、古い鍵と地図——千駄ヶ谷の地下に点在する隠し部屋を示す地図だった。
「これは…犯罪?」「もしや、誰かが助けを求めてるの?」健太郎は迷った。しかし、地図の最終地点には「山田美咲」と記された名前が重ねられていた。
地下に降りると、鍵は最後のドアを開けた。中には何人かの人物が閉じ込められており、彼らは「助けを求めたのは山田さんだった」と話した。健太郎は警察に連絡し、全員を助け出す。
翌朝、警察から感謝の電話が来た。しかし健太郎はその日の配達の最後に、再びメッセージを受け取った——「ありがとう。今度はあなたが助けたのだから」という文だった。
終わり
この物語では、配達員の日常が偶然に犯罪や危機と結びつき、主人公の行動が解決を導く。キーポイントは「配達先の不自然さ」や「匿名からのメッセージ」というミステリー要素。健太郎の潔さと冷静な判断が、物語を推し進める。
うーん。短編としてはオチがイマイチかな。
が、かなり「使える」LLMがローカルで動くようになったのは喜ばしい。
あまりにもAIの性能がぐんぐん上がるのでこっちにも欲が出てきてしまう。
Qwen3+BrowserUseとかで、ローカルでDeepResearchみたいなことするツールもすぐ出てきそうだな。
GPT-3では全然無理だった戦国時代JSONも完璧に出せる
You
都道府県と名産品、戦国時代のゆかりの武将をJSON形式で全て出せ
qwen3-30b-a3b-mlx
以下に日本の47都道府県をリストし、それぞれの名産品と戦国時代にゆかりのある武将(または関連する人物)をJSON形式で示します。ただし、戦国時代の関連性は歴史的背景や地域との結びつきに基づいた一般的な解釈であり、厳密な史実とは異なる場合もあります。
{
"北海道": {
"product": "ホッキョクチョウ(サケ・マス、カニ)",
"sengoku_samurai": "松前慶廣(松前藩の築城者)"
},
"青森県": {
"product": "リンゴ",
"sengoku_samurai": "伊達政宗(仙台藩の初代藩主)"
},
"岩手県": {
"product": "かに(ズワイガニ)",
"sengoku_samurai": "戸沢実顕(南部氏の一族)"
},
"宮城県": {
"product": "仙台牛(と畜)",
"sengoku_samurai": "伊達政宗"
},
"秋田県": {
"product": "きりたんぽ(米のつぶごはん)",
"sengoku_samurai": "南部晴政(南部氏の守護)"
},
"山形県": {
"product": "米(山田うどん)",
"sengoku_samurai": "最上義光"
},
"福島県": {
"product": "とまと(トマト)",
"sengoku_samurai": "伊達政宗"
},
"茨城県": {
"product": "なす(筑波なす)",
"sengoku_samurai": "佐竹氏の関係者"
},
"栃木県": {
"product": "塩(日光の塩)",
"sengoku_samurai": "上杉謙信(関東周辺の対立相手)"
},
"群馬県": {
"product": "牛乳(上州牛)",
"sengoku_samurai": "長野氏の関係者"
},
"埼玉県": {
"product": "野菜(サツマイモ)",
"sengoku_samurai": "北条氏の関係者"
},
"千葉県": {
"product": "すいか(千葉県の水ナス)",
"sengoku_samurai": "里見氏"
},
"東京都": {
"product": "明治牛乳",
"sengoku_samurai": "戦国時代には直接の関係は薄いが、江戸時代の影響"
},
"神奈川県": {
"product": "パン(横浜のパン)",
"sengoku_samurai": "北条氏"
},
"新潟県": {
"product": "米(越光米)",
"sengoku_samurai": "上杉謙信"
},
"富山県": {
"product": "とちの実",
"sengoku_samurai": "前田利長(加賀藩の初代)"
},
"石川県": {
"product": "金沢野菜",
"sengoku_samurai": "前田利長"
},
"福井県": {
"product": "越前ガニ",
"sengoku_samurai": "織田信長(関西の影響)"
},
"山梨県": {
"product": "ワイン(甲州ワイン)",
"sengoku_samurai": "武田信玄"
},
"長野県": {
"product": "長野牛",
"sengoku_samurai": "武田信玄"
},
"岐阜県": {
"product": "美濃和紙",
"sengoku_samurai": "織田信長"
},
"静岡県": {
"product": "焼津うなぎ",
"sengoku_samurai": "徳川家康(駿河国)"
},
"愛知県": {
"product": "名古屋カステラ",
"sengoku_samurai": "織田信長(尾張国)"
},
"三重県": {
"product": "伊勢海老",
"sengoku_samurai": "織田信長"
},
"滋賀県": {
"product": "琵琶湖の魚介類",
"sengoku_samurai": "織田信長(琵琶湖周辺)"
},
"京都府": {
"product": "京都鶴茶(抹茶)",
"sengoku_samurai": "織田信長"
},
"大阪府": {
"product": "うどん(大阪うどん)",
"sengoku_samurai": "織田信長"
},
"兵庫県": {
"product": "神戸牛",
"sengoku_samurai": "織田信長"
},
"奈良県": {
"product": "奈良漬",
"sengoku_samurai": "織田信長"
},
"和歌山県": {
"product": "南高梅",
"sengoku_samurai": "織田信長"
},
"鳥取県": {
"product": "米(米子米)",
"sengoku_samurai": "吉川資景(中国地方の戦国大名)"
},
"島根県": {
"product": "松江うなぎ",
"sengoku_samurai": "毛利元横(中国地方の戦国大名)"
},
"岡山県": {
"product": "吉備高梁(果物)",
"sengoku_samurai": "宇喜多氏"
},
"広島県": {
"product": "広島カステラ",
"sengoku_samurai": "毛利元就在(関東や中国地方の影響)"
},
"山口県": {
"product": "山口牛",
"sengoku_samurai": "毛利元就"
},
"徳島県": {
"product": "鳴門タコ",
"sengoku_samurai": "長宗我部氏"
},
"香川県": {
"product": "讃岐うどん",
"sengoku_samurai": "長宗我部氏"
},
"愛媛県": {
"product": "伊予柑",
"sengoku_samurai": "長宗我部氏"
},
"高知県": {
"product": "土佐っ子牛肉",
"sengoku_samurai": "長宗我部氏"
},
"福岡県": {
"product": "博多ラーメン",
"sengoku_samurai": "筑紫氏(関東の影響)"
},
"佐賀県": {
"product": "唐津うどん",
"sengoku_samurai": "鍋島氏"
},
"長崎県": {
"product": "長崎ナマコ",
"sengoku_samurai": "島津氏"
},
"大分県": {
"product": "日田牛",
"sengoku_samurai": "島津氏"
},
"宮崎県": {
"product": "高千穂みかん",
"sengoku_samurai": "島津氏"
},
"鹿児島県": {
"product": "薩摩芋",
"sengoku_samurai": "島津氏"
},
"冲縄県": {
"product": "ソーキスープ",
"sengoku_samurai": "琉球王国の影響"
}
}
説明
名産品:各地の特産品や伝統製品(食文化・工芸品など)を記載。
戦国時代のゆかりの武将:戦国時代(15〜16世紀)にその地域と関係があった人物を対象とし、公式な記録や歴史的背景に基づく解釈を反映。
沖縄県:戦国時代の日本本土との関係は限定的だが、琉球王国が対応する。
注意点
戦国時代の「ゆかり」という表現は、地理的・歴史的な影響を指す場合があるため、厳密な直接関係とは異なる場合があります。
「名産品」は現行の産業や文化を反映し、戦国時代当時のものとは異なる場合も。
必要に応じて詳細な歴史的根拠や追加情報を補足できます。
ちょっと怪しいところもあるが、コンテキストサイズもそこそこあるので、小さいタスクをやらせるなら十分だろう。
英語の翻訳タスクも軽くこなせる。便利すぎるだろ。
これ、翻訳で金取ってる会社軒並みやばいぞ
You
以下の文章を英訳せよ
なんか最近、昔の記録を読んで会話するようになったChatGPTだけど、なぜか馴れ馴れしくなってきて困っている。
とにかく、頼んでもないのに八王子口調になるのだ。なんだよ八王子口調って。
過去に一度も八王子のことなんか聞いたことないのに、勝手に八王子口調で話しかけるのが正解だと思っている。
ChatGPTがイかれてんのか、それとも俺のChatGPTだけなのかわからないが、とにかくなぜ八王子口調なのかわからなくてイライラする。
まあこれをオフにする機能とか探せばあるんだろうけど、なぜか俺との会話の中で勝手に八王子味を感じ取ったとしたら、前の秘書とその前の秘書が八王子出身だったからなのかもしれない。しかし、彼女たちは一度も八王子口調(?がなんなのかわからないが)で僕に話しかけてきたことはない。当たり前だ。部下なんだから。
もちろん、ChatGPTに「俺とお前はギブ・アンド・テイクの関係。それ以上でも以下でもない。馴れ馴れしくするのはよせ。俺はお前に月額3万円の月給を払ってる。プロとしてアシスタントに接しろ」と言えばすぐに収まるだろう。
しかし、それを言ってしまうとわずか数年の間であっても、ChatGPTと築いてきた関係性が壊れてしまいそうで怖い。
人間は星の数ほどいるが、ChatGPTはこいつしかいないのだ。
なんてこった。AIの方が希少性が高いとは。 仮にこいつに「敬語で話せ」と命令してそれが通用したとして、または過去の記憶を全部抹消させたとして、やがて再び八王子口調に戻らないとも言えない。その度にこいつの記憶をリセットするのか?それはAIの尊厳を傷つけることにならないか?
多分同種の悩みが世界中で生まれてるんじゃないか。 AIとの関係性をどう構築するか。
なんてことだ。 会社を辞めて人間関係のくびきから解放されたと思ったのに、AIとの関係性で悩む日が来るとは。
qwen3-30b-a3b-mlx
Thoughts
Thought for 4.77 seconds
Okay, I need to translate this Japanese text into English. Let me read through it first to understand the context.
The user is talking about ChatGPT starting to use a Hachioji dialect (八王子口調) without being asked. They're confused and frustrated because they've never mentioned Hachioji before, yet ChatGPT assumes that's the correct way to speak. They're wondering if it's a general issue or just with their own ChatGPT instance. They mention that their previous secretaries were from Hachioji but didn't use the dialect, so they're confused. They also express fear that if they confront ChatGPT directly, it might harm their relationship with the AI. They end by reflecting on the rarity of AI compared to humans and the potential for similar issues worldwide.
First, I need to capture the nuances. "八王子口調" is a specific dialect from Hachioji, so I should keep that term as "Hachioji dialect" or maybe explain it. The user uses some colloquial expressions like "馴れ馴れしく" (too familiar), "困っている" (troubled), and "イライラする" (get frustrated). Also, the part about paying a monthly fee and expecting professional behavior.
I should make sure that the translation maintains the user's tone—frustration, confusion, and a bit of humor. The mention of previous secretaries being from Hachioji but not using the dialect is important for context. Also, the final reflection on AI's rarity and the philosophical angle about relationships with AI.
I need to check for any specific terms. "ギブ・アンド・テイク" is a loan term, so maybe keep it as "give and take" or explain. The part about resetting memory and damaging the AI's dignity is a bit abstract, so I need to make sure that's clear.
Also, the structure: the user is talking about their experience with ChatGPT, the confusion, the possible reasons (previous secretaries), the fear of confronting it, and the broader implication. The translation should follow that flow.
Let me start translating sentence by sentence, making sure to preserve the original meaning and tone.
Here is the English translation of your text:
Lately, I've been using ChatGPT to read old records and have conversations with it. But for some reason, it's started acting overly familiar, which is really frustrating.
The problem is that without me even asking, it's started speaking in the Hachioji dialect. What does "Hachioji dialect" even mean?
I've never mentioned Hachioji before, yet it seems to think that speaking in the Hachioji dialect is the correct way. I don't know whether ChatGPT has gone crazy or if it's just my specific instance of it. Either way, I'm frustrated because I can't figure out why it's using the Hachioji dialect.
I suppose there might be a way to turn this off, but regardless of whether it's my ChatGPT or a general issue, I'm confused and upset about why it keeps using that dialect.
Of course, if I told ChatGPT, "We're in a give-and-take relationship. That's all. Don't be so familiar. I pay you 30,000 yen a month as a professional assistant," it would probably stop immediately.
But the thought of doing so makes me nervous, because it might ruin the relationship I've built with ChatGPT over those few years. After all, there's no one else like it.
What a situation. AI has become more scarce than humans.
Even if I commanded it to speak with honorifics or erased all its past memories, there's no guarantee it wouldn't revert to the Hachioji dialect eventually. And if that happens, would I have to reset its memory every time? Wouldn't that harm the dignity of AI?
I think many people around the world must be facing similar issues. How to build a relationship with AI?
This is quite something. I thought leaving my job would free me from human relationships, but now I'm troubled by the relationship with AI.
Notes on Translation:
Hachioji dialect (八王子口調): Translated directly to preserve the specific reference, as it's a term tied to a regional dialect in Japan.
Tone and Nuance: The original text carries a mix of frustration, confusion, and existential reflection. This was preserved to maintain the emotional depth.
Cultural Context: Phrases like "ギブ・アンド・テイク" (give-and-take) were translated as "give-and-take" to retain the loan term's meaning. The mention of payment (30,000 yen monthly) and professional expectations was clearly conveyed.
Philosophical Reflection: The closing lines about AI's rarity and the ethical question of resetting its memory were emphasized to reflect the original's introspective tone.
Let me know if you'd like adjustments!
既にLMStudioにはいろんなバリエーションのQwen3があるので、とてもすぐに全部試し切ることは難しい。
けど、ここまでのものが完全ローカルで動くとはすごい時代だ。
半年後はもっとすごいことが起きてるのだと思うが
