FIELD GUIDE / 01
Opening SIGNAL / 読み込んでいます
FIELD GUIDE / 01
New capabilities, access and limitations. Read past the leaderboard.新しい能力、使える範囲、残る制約。ランキングの先まで読む。
SIGNAL reading guide · SIGNALの読み方ガイド ·
A release can change what a model does, who can use it, or what it costs. Those are separate questions. We follow model announcements with that distinction in mind: a downloadable model is not the same offer as a hosted API, and a provider’s benchmark claim is not a result from your own work. Start with the concrete change in the brief, then use the source links to check the version and conditions that matter to you.新モデルの発表では「何ができるか」「誰が使えるか」「いくらかかるか」を分けて読みます。ダウンロードできるモデルと、API経由で使うサービスは同じ提供形態ではありません。また、提供元のベンチマーク結果は、自分の作業でも同じ成果が出ることを示すものではありません。ここでは発表を時系列で追い、記事の要点から具体的な変化をつかみ、原典で対象のバージョンや利用条件を確認できるようにしています。
01
Which task changed: writing, reasoning, images, audio or tool use? Is the claim tied to a specific version?文章、推論、画像、音声、ツール操作。どの作業が変わり、どのバージョンの話なのか。
02
Is this an announcement, a limited preview or something available now? Check the region, plan and API or download link.発表段階、限定プレビュー、一般提供のどれか。地域、プラン、APIや配布先を確認する。
03
Who ran the evaluation, under which conditions? Keep the source’s limitations attached to the result.誰が、どんな条件で評価したか。結果だけを抜き出さず、原典の制約と一緒に読む。
These are reading prompts, not findings about every story below.以下の個別記事に対する評価ではなく、読むときの確認ポイントです。
THE RUNNING STORY
Matching stories published in the past 30 days, newest first. Dates below show when SIGNAL added each brief. Follow-ups and unverified reports remain visible.過去30日間の関連記事を新しい順に表示します。日付はSIGNALの掲載日です。続報や、区別して表示した未確認情報も含みます。
Topic RSS contains up to 100 standard reports from this window. Unverified reports are excluded.テーマRSSはこの期間の通常記事を最大100本配信します。未確認情報は含みません。
← All topics← ほかのテーマを見るAlibaba CEO Eddie Wu announced plans to train a 5-10 trillion parameter AI model and unveiled a new Zhenwu V900 AI chip at Apsara Conference.AlibabaのCEO Eddie Wuが、5兆から10兆パラメータのAIモデル訓練計画と新型AIチップZhenwu V900を発表した。
Xiaomi has released Pro and Flash model weights, with image, video and text capabilities. API and desktop access are available; open weights do not mean a free API.XiaomiがProとFlashの重みを公開。画像・動画・テキストを扱い、APIやデスクトップでも利用できます。重みの公開とAPIの無料提供は別です。
Higgsfield AI says its engineers now ship new video ad creation features within a day using OpenAI's GPT-6 Astra model.Higgsfield AIはOpenAIのGPT-6 Astraを使い、新しい動画広告作成機能をわずか1日で開発できるようになったと発表した。
SpaceXAI launched Grok 4.7, a new coding and knowledge-work model priced at $2 per million input and $6 per million output tokens.SpaceXAIは、コーディングと知識労働向けの新モデルGrok 4.7を公開しました。料金は入力100万トークンあたり2ドル、出力100万トークンあたり6ドルです。
The creator describes an experimental language model that trains on an 8GB GPU. It is a small prototype; weights are not yet published.開発者が、8GBのGPUで学習する言語モデルの試作を紹介しました。作者自身が小規模な実験と位置づけており、モデルの重みはまだ公開されていません。
A blog describes GPT-6 Astra decoding a German wartime message. The account is a reported case study, not an independent evaluation of the model.ブログで、GPT-6 Astraを使って第一次大戦中のドイツ軍の暗号を解読した事例が紹介されました。モデルの能力を独立に検証した評価ではありません。
SpaceXAI released Grok Voice Transcribe 2.0, a speech-to-text model it says is twice as accurate as version 1.0 at the same price.SpaceXAIが音声認識モデルGrok Voice Transcribe 2.0をリリースしたと報じられました。精度が従来の2倍で価格は据え置き、という評価は同社によるものです。
Users spotted signs of Anthropic secretly testing Claude Fable 5.2 in the wild, generating advanced coded animations, with Opus 5.5 rumored to launch this week.AnthropicはClaude Fable 5.2とOpus 5.5をリリース前にテストしており、Fable 5.2は外部アセットなしでJavaScriptのみで対話的なアニメーションを生成できる能力を示している。
StepFun unveiled Step 5 Preview, a sparse Mixture-of-Experts model with 600B total parameters and 27B active per token, to be released with open weights on October 15.Stepfunが「Step 5 Preview」を発表。600Bのパラメータを持ち、1M トークンのコンテキストウィンドウとビジョン入力に対応し、10月15日にオープンウェイトで公開予定。
Qwen-Image-2.1 combines image generation and editing. Its visual generation component has 7 billion parameters; comparative performance claims come from Qwen’s own benchmarks.アリババのQwenチームが、画像生成と編集を統合したモデルを公開しました。70億パラメータは画像生成部分の規模です。他モデルを上回るという評価は同社のベンチマークに基づく主張です。