{"id":9915,"date":"2026-10-05T09:29:52","date_gmt":"2026-10-05T00:29:52","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/10\/05\/local-ai-weekly-2026-10-week-1\/"},"modified":"2026-10-05T09:29:52","modified_gmt":"2026-10-05T00:29:52","slug":"local-ai-weekly-2026-10-week-1","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/10\/05\/local-ai-weekly-2026-10-week-1\/","title":{"rendered":"Local AI Model Weekly Highlights: Structured Output & Video LoRAs"},"content":{"rendered":"

<\/p>\n

This Week in Numbers<\/h2>\n

Counted by Local Model Watch from the articles published this week.<\/em><\/p>\n

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Item<\/th>\nCount<\/th>\n<\/tr>\n<\/thead>\n
Articles published<\/td>\n36<\/td>\n<\/tr>\n
\u3000Image, Video and Audio<\/td>\n13<\/td>\n<\/tr>\n
\u3000New Models<\/td>\n12<\/td>\n<\/tr>\n
\u3000Engines and Tools<\/td>\n10<\/td>\n<\/tr>\n
\u3000Community<\/td>\n1<\/td>\n<\/tr>\n
New models covered<\/td>\n25<\/td>\n<\/tr>\n
\u3000fit in 8GB of VRAM (est.)<\/td>\n5<\/td>\n<\/tr>\n
\u3000fit in 12GB of VRAM (est.)<\/td>\n7<\/td>\n<\/tr>\n
\u3000fit in 16GB of VRAM (est.)<\/td>\n8<\/td>\n<\/tr>\n
\u3000fit in 24GB of VRAM (est.)<\/td>\n11<\/td>\n<\/tr>\n
\u3000parameters 15\u201340B<\/td>\n6<\/td>\n<\/tr>\n
\u3000parameters 4\u201315B<\/td>\n5<\/td>\n<\/tr>\n
\u3000parameters up to 4B<\/td>\n2<\/td>\n<\/tr>\n
\u3000parameters over 40B<\/td>\n2<\/td>\n<\/tr>\n
Converted builds appended to earlier articles<\/td>\n18 (GGUF 10, FP8 1, MLX 7)<\/td>\n<\/tr>\n
Most active publishers<\/td>\nefficient-large-model (5), ggml-org (3), unslothai\/unsloth (2)<\/td>\n<\/tr>\n
Articles still marked unverified<\/td>\n1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Trending models we did not cover separately<\/h3>\n

Hugging Face repositories that trended this week but did not get their own article. Listed for reference; we have not reviewed their model cards.<\/em><\/p>\n

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Model<\/th>\nLikes<\/th>\nDownloads<\/th>\nWhy no article<\/th>\n<\/tr>\n<\/thead>\n
PSRben\/VisionHOPE<\/a><\/td>\n402<\/td>\n1,516<\/td>\npublisher not on our notable list<\/td>\n<\/tr>\n
Aleph-Alpha\/Kolibri-1<\/a><\/td>\n388<\/td>\n1,135<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n
ISTA-DASLab\/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF<\/a><\/td>\n260<\/td>\n351,230<\/td>\npublisher not on our notable list<\/td>\n<\/tr>\n
NaiveAI\/Naive-N0.5-Flash<\/a><\/td>\n165<\/td>\n1,920<\/td>\npublisher not on our notable list<\/td>\n<\/tr>\n
bartowski\/FrogNano-4B-2609-GGUF<\/a><\/td>\n11<\/td>\n7,519<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n
bartowski\/bytkim_Qwen3.8-27B-pi-GGUF<\/a><\/td>\n7<\/td>\n3,909<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n
bartowski\/GLM-5.3-Flash-BF16-GGUF<\/a><\/td>\n4<\/td>\n2,288<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n
bartowski\/cosmicoptima_computer-10-GGUF<\/a><\/td>\n3<\/td>\n6,853<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n
bartowski\/OmniJev_OneJev-9B-GGUF<\/a><\/td>\n3<\/td>\n3,333<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n
bartowski\/Lythri_Lythri-4B-A2B-GGUF<\/a><\/td>\n2<\/td>\n2,819<\/td>\nconverted build without a parent article<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Patch releases we did not cover separately<\/h3>\n

Releases of watched projects that were patch-level or had short notes. Each project’s page lists every version.<\/em><\/p>\n

\n\n\n\n\n\n\n\n\n\n
Project<\/th>\nVersion<\/th>\nRelease notes<\/th>\n<\/tr>\n<\/thead>\n
magnitudedev\/magnitude<\/td>\n@magnitudedev\/cli@0.2.2<\/td>\nGitHub<\/a><\/td>\n<\/tr>\n
magnitudedev\/magnitude<\/td>\n@magnitudedev\/cli@0.2.3<\/td>\nGitHub<\/a><\/td>\n<\/tr>\n
ollama\/ollama<\/td>\nv0.35.1<\/td>\nGitHub<\/a><\/td>\n<\/tr>\n
turboderp-org\/exllamav3<\/td>\nv1.5.4<\/td>\nGitHub<\/a><\/td>\n<\/tr>\n
unslothai\/unsloth<\/td>\nv0.1.902-beta<\/td>\nGitHub<\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Models Gaining the Most Likes<\/h3>\n

Compiled by Local Model Watch from weekly snapshots of models we’ve covered. Likes are cumulative on Hugging Face; downloads are its trailing-30-day count.<\/em><\/p>\n

\n

\u2192 Scroll horizontally to see all columns<\/p>\n\n\n\n\n\n\n\n\n\n
Model<\/th>\nLikes gained<\/th>\nTotal likes<\/th>\nDownloads (30d)<\/th>\nPeriod<\/th>\n<\/tr>\n<\/thead>\n
Edge0\/Audio8-ASR-Infinite<\/a><\/td>\n+1,401<\/td>\n2,426<\/td>\n40,004<\/td>\n2026-W40 \u2192 2026-W41<\/td>\n<\/tr>\n
Qwen\/Qwen3.8-27B<\/a><\/td>\n+515<\/td>\n16,939<\/td>\n6,821,761<\/td>\n2026-W40 \u2192 2026-W41<\/td>\n<\/tr>\n
deepseek-ai\/DeepSeek-V4.1-Flash<\/a><\/td>\n+283<\/td>\n4,093<\/td>\n798,422<\/td>\n2026-W40 \u2192 2026-W41<\/td>\n<\/tr>\n
Viggle\/Qwen-Image-2.1-viggle-turbo<\/a><\/td>\n+246<\/td>\n587<\/td>\n272,896<\/td>\n2026-W40 \u2192 2026-W41<\/td>\n<\/tr>\n
prism-ml\/Ternary-Bonsai-2-27B-gguf<\/a><\/td>\n+228<\/td>\n2,416<\/td>\n4,045,810<\/td>\n2026-W40 \u2192 2026-W41<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

<\/p>\n

Highlights of the Week<\/h2>\n

The most notable highlights surrounding local AI models this week are the following three points.<\/p>\n

First, there is the rise of models specialized in structured output and decision-making, along with advances in their GGUF support. Models optimized for specific tasks and structured data processing, such as “clef<\/a>” and “clef-flash<\/a>” released by Cloudflare, and “OpenJev-GGUF<\/a>” released by ggml-org, have appeared one after another. This significantly improves the practicality of agents and automation tools that can be run on local PCs.<\/p>\n

Second, LoRA adapters for video generation models have been released very actively. Methods for practically customizing and controlling existing powerful models have been enriched, such as the “LongLive-Plug” series by Efficient-Large-Model and LoRAs for video inpainting and quality conversion by Lightricks.<\/p>\n

Third, competition to accelerate local inference engines and expand their features is intensifying. The emergence of the new Rust-based engine “Magnitude<\/a>” and the addition of new APIs for decision-making models in “Ollama v0.35.0<\/a>” are rapidly establishing the infrastructure for developers to run models more comfortably in their local environments.<\/p>\n

Trends by Category<\/h2>\n

Text Generation<\/h3>\n

This week saw the release of numerous practical models specialized for specific use cases and reasoning capabilities. Optimization is progressing not only for general conversation, but also for tasks requiring structured data output and complex thought processes.<\/p>\n

Cloudflare released “clef<\/a>” (27.4B) and its lightweight version “clef-flash<\/a>” (9.4B), which are specialized for structured decision-making tasks. These exhibit high accuracy in use cases such as API integration and data extraction. Additionally, domestic Japanese developer ELYZA released the 32B and 33B models of “ELYZA-Thinking-1.0<\/a>“, an inference model capable of outputting thought processes, enabling local execution of advanced reasoning tasks in Japanese environments. Furthermore, previously released models like “Xing4.0-29B-A4B<\/a>` and “Hemmingway-1<\/a>” have also been covered on this site, further enriching the options in the mid-size tier.<\/p>\n

Image, Video, and Audio<\/h3>\n

It was a week notable for advancements in video generation control technology and the release of unique models specialized in audio processing.<\/p>\n

In the field of video generation, numerous LoRAs were released to efficiently control underlying large-scale models. In particular, LoRA adapters targeting Wan2.1 and Wan2.2, such as “LongLive-Plug-Wan2.2-TI2V-5B-cfg<\/a>“, enable high-quality generation with fewer steps and camera angle control. In the audio field, Google released “DiarizationLM-Gemma-4-E4B-v1<\/a>“, which is suitable for speech transcription and speaker diarization, supporting the local execution of practical audio processing. Note that the release of the Turkish synthetic speech dataset “alania-synthetic-speech-tr<\/a>” has also been reported, but please be aware that this is unverified information that has not been officially confirmed.<\/p>\n

Engines and Tools<\/h3>\n

Development competition continues with the goal of maximizing execution speed in local environments and rapidly supporting the latest models.<\/p>\n

The image generation GUI “ComfyUI v0.38.0<\/a>” was released, adding support for the latest Hunyuan Image 3.5 and Qwen-Image 2.1. In addition, “Ollama v0.35.0<\/a>“, a popular local LLM execution environment, implemented a new API to make decision-making models easier to handle. Furthermore, a new inference engine written in Rust called “Magnitude<\/a>” appeared, followed by rapid-fire updates such as “Magnitude CLI 0.2.4<\/a>” and “@magnitudedev\/cli v0.2.5<\/a>” which advance memory reduction and Apple Silicon optimization. The evolution of these tools is building an environment where the latest models can be run efficiently even with limited hardware resources.<\/p>\n

Industry News<\/h2>\n

These are announcements from companies and research institutions that were not made into standalone articles because they are not topics about running locally. Only the key points are listed.<\/p>\n