Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions strataqwen38swift/Chart.yaml
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
apiVersion: v2
appVersion: 0.2.11
appVersion: 0.2.12
description: Qwen3.8-Flash-Next GSQ-RCO IQ3_S with Strata and Model Console
name: strataqwen38swift
type: application
version: 0.0.1
version: 1.0.3
4 changes: 3 additions & 1 deletion strataqwen38swift/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
appid: strataqwen38swift
title: Qwen3.8 Swift 1.5 IQ3_S
version: 0.0.1
version: 1.0.3
categories:
- AI
- models
Expand Down Expand Up @@ -42,6 +42,8 @@ workloadReplicas:
spec:
onlyAdmin: true
versionName: Swift 1.5 GSQ-RCO IQ3_S
upgradeDescription: |
- Strata 0.2.12 (engine 0.1.58): the DFlash2 drafting helper now starts by default on this app, making generation faster; fixes for tool calls in several API clients (Claude Desktop model name, tool-call tag forms, MCP name matching, forced tool_choice); corrected server speed reports; and a more stable engine on long CUDA prompts. Model, 262K context and image input are unchanged.
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 is UkisAI's fine-tune of Alibaba's Qwen3.8-Flash-Next, built to answer sooner: the authors report 63 % fewer thinking tokens and answers starting 1.8x earlier than the original model. This app runs its GSQ-RCO IQ3_S quantization locally using Strata.

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/i18n/de-DE/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
Erstinstallationen sind robuster: Hat ein Systemdienst das Modell-Cache-Verzeichnis im gemeinsamen Hugging-Face-Cache vorab angelegt, übernimmt die App es vor dem Herunterladen zurück; der Download beim ersten Start endet nicht mehr mit einem Berechtigungsfehler.
- Strata 0.2.12 (Engine 0.1.58): Der DFlash2-Entwurfshelfer startet bei dieser App jetzt standardmäßig, wodurch die Generierung schneller wird; Korrekturen an Tool-Aufrufen in mehreren API-Clients (Claude-Desktop-Modellname, Formen der Tool-Call-Tags, MCP-Namensabgleich, erzwungene tool_choice); korrigierte Geschwindigkeitsanzeigen des Servers; eine stabilere Engine bei langen CUDA-Prompts. Modell, 262K-Kontext und Bildeingabe bleiben unverändert.
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 ist ein Feintuning von UkisAI auf Basis von Alibabas Qwen3.8-Flash-Next, gebaut, um früher zu antworten: Die Autoren berichten über 63 % weniger Thinking-Tokens und Antworten, die 1,8x früher beginnen als beim Originalmodell. Diese App führt die GSQ-RCO-IQ3_S-Quantisierung lokal mit Strata aus.

Expand Down
2 changes: 2 additions & 0 deletions strataqwen38swift/i18n/en-US/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,8 @@ metadata:
title: Qwen3.8 Swift 1.5 IQ3_S
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
- Strata 0.2.12 (engine 0.1.58): the DFlash2 drafting helper now starts by default on this app, making generation faster; fixes for tool calls in several API clients (Claude Desktop model name, tool-call tag forms, MCP name matching, forced tool_choice); corrected server speed reports; and a more stable engine on long CUDA prompts. Model, 262K context and image input are unchanged.
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 is UkisAI's fine-tune of Alibaba's Qwen3.8-Flash-Next, built to answer sooner: the authors report 63 % fewer thinking tokens and answers starting 1.8x earlier than the original model. This app runs its GSQ-RCO IQ3_S quantization locally using Strata.

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/i18n/es-ES/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
Las primeras instalaciones son más sólidas: si un servicio del sistema ha creado previamente el directorio de caché del modelo en la caché compartida de Hugging Face, la app recupera ese directorio antes de descargar; la descarga inicial ya no se detiene por un error de permisos.
- Strata 0.2.12 (motor 0.1.58): el asistente de borrador DFlash2 ahora se inicia de forma predeterminada en esta app, lo que acelera la generación; correcciones de llamadas a herramientas en varios clientes de API (nombre de modelo de Claude Desktop, formas de etiquetas de llamadas a herramientas, coincidencia de nombres MCP, tool_choice forzado); informes de velocidad del servidor corregidos; y un motor más estable con prompts CUDA largos. El modelo, el contexto de 262K y la entrada de imágenes no cambian.
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 es un fine-tuning de UkisAI sobre Qwen3.8-Flash-Next de Alibaba, creado para responder antes: los autores reportan un 63 % menos de tokens de razonamiento y respuestas que comienzan 1,8x antes que el modelo original. Esta app ejecuta localmente su cuantización GSQ-RCO IQ3_S mediante Strata.

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/i18n/fr-FR/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
Les premières installations sont plus fiables : si un service système a précréé le répertoire de cache du modèle dans le cache Hugging Face partagé, l'application le reprend avant le téléchargement ; le téléchargement initial ne s'arrête plus sur une erreur de permissions.
- Strata 0.2.12 (moteur 0.1.58) : l'assistant d'ébauche DFlash2 démarre désormais par défaut sur cette application et accélère la génération ; corrections des appels d'outils dans plusieurs clients d'API (nom de modèle Claude Desktop, formes de balises d'appel d'outils, correspondance des noms MCP, tool_choice forcé) ; rapports de vitesse du serveur corrigés ; moteur plus stable sur les longs prompts CUDA. Le modèle, le contexte 262K et l'entrée d'images restent inchangés.
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 est un fine-tuning d'UkisAI du Qwen3.8-Flash-Next d'Alibaba, conçu pour répondre plus tôt : les auteurs rapportent 63 % de tokens de raisonnement en moins et des réponses démarrant 1,8x plus tôt que le modèle d'origine. Cette application exécute localement sa quantification GSQ-RCO IQ3_S via Strata.

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/i18n/it-IT/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
Le prime installazioni sono più robuste: se un servizio di sistema ha precreato la directory di cache del modello nella cache condivisa di Hugging Face, l'app la riprende prima di scaricare; il download iniziale non si ferma più per un errore di permessi.
- Strata 0.2.12 (motore 0.1.58): l'assistente di bozza DFlash2 ora si avvia per impostazione predefinita in questa app, rendendo la generazione più rapida; correzioni delle chiamate a strumenti in diversi client API (nome del modello di Claude Desktop, forme dei tag di chiamata a strumento, corrispondenza dei nomi MCP, tool_choice forzato); rapporti di velocità del server corretti; e un motore più stabile con prompt CUDA lunghi. Modello, contesto 262K e input di immagini restano invariati.
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 è un fine-tuning di UkisAI di Qwen3.8-Flash-Next di Alibaba, creato per rispondere prima: gli autori riportano il 63 % di token di ragionamento in meno e risposte che iniziano 1,8x prima rispetto al modello originale. L'app esegue localmente la quantizzazione GSQ-RCO IQ3_S tramite Strata.

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/i18n/ja-JP/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
初回インストールが堅牢になりました: システム サービスが共有 Hugging Face キャッシュにモデルのキャッシュ ディレクトリを事前作成している場合、アプリはダウンロード前にその所有権を取り戻すため、初回のダウンロードが権限エラーで停止しなくなりました。
- Strata 0.2.12(エンジン0.1.58): このアプリではDFlash2ドラフターが既定で起動するようになり、生成が速くなりました。複数のAPIクライアントでのツール呼び出し(Claude Desktopのモデル名、ツール呼び出しタグの形式、MCP名の末尾一致、tool_choiceの強制)の修正、サーバーの速度表示の訂正、長いCUDAプロンプトでの安定性改善も含みます。モデル、262Kコンテキスト、画像入力は変更ありません。
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 は、Alibaba の Qwen3.8-Flash-Next を UkisAI がファインチューニングしたモデルで、より早く回答することを目指しています。作者の報告では、思考トークンが 63% 減少し、元のモデルより 1.8 倍早く回答が始まります。このアプリは、その GSQ-RCO IQ3_S 量子化版を Strata でローカル実行します。

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/i18n/zh-CN/OlaresManifest.yaml
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ metadata:
description: "Qwen3.8-Flash-Next Swift 1.5 for text generation, reasoning, and tool calling, powered locally by Strata"
spec:
upgradeDescription: |
首次安装更加稳妥:当系统服务预先创建了模型在共享 Hugging Face 缓存中的目录时,应用会在下载前收回该目录的所有权,首次运行的下载不再因权限错误而停止。
- Strata 0.2.12(引擎 0.1.58):DFlash2 草稿模型现在在本应用中默认启动,生成更快;修复多个 API 客户端的工具调用(Claude Desktop 模型名、工具调用标签写法、MCP 名称匹配、强制 tool_choice);更正服务器速度报告;改进长 CUDA 提示词时的引擎稳定性。模型、262K 上下文与图像输入保持不变。
fullDescription: |
Qwen3.8-Flash-Next Swift 1.5 是 UkisAI 在阿里巴巴 Qwen3.8-Flash-Next 基础上微调的模型,目标是更早给出回答:作者报告其思考 token 减少 63%,回答起点比原模型提前 1.8 倍。本应用通过 Strata 在本地运行其 GSQ-RCO IQ3_S 量化版本。

Expand Down
2 changes: 1 addition & 1 deletion strataqwen38swift/values.yaml
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
images:
strata: docker.io/mushrxxm/strata:0.2.11-cuda13-amd64-repo
strata: docker.io/mushrxxm/strata:0.2.12-cuda13-amd64-repo
llmInit: docker.io/beclab/llm-init:v1.7.24
workloads:
llminit:
Expand Down
Loading