Every week, I organize 400 to 700 links into roughly 60 categories as part of my ongoing effort to learn about AI. This is my personal notebook, which I enjoy sharing with friends… a hobby and a labor of love, rather than a commercial publication or product.
If you arrived here through a search or shared link, this page collects the links I found for Locally Run for the week ending August 21, 2026.
As part of my learning process, I like to automate the category covers. It gives me a chance to learn Python and APIs.
This week’s cover prompt was written using Claude Opus 4.7, and the image was generated using Nano Banana 2.1.
Category Cover Image Prompt: Bright sunlit architectural photograph of a 1929 Chicago Art Deco limestone skyscraper with setbacks and dark steel windows, where a single stylized Art Deco herm emerges from a pier above the main entrance and cradles a small carved limestone bungalow against its chest with tiny geometric circuit reliefs for windows, the word LOCAL incised in large Deco capitals on the limestone lintel above the doorway, warm honey-cream Indiana limestone against a deep blue sky with crisp raking shadows, neighboring Deco buildings receding behind, perspective-corrected verticals, realistic architecture lens.
This Week in Locally Run News
Summary by Claude Sonnet 5.5 based on this week’s links:
- Qwen3.8-27B took over the week: Cline said it became the top local model in its tool after four days, and also claimed it scored at frontier level on the Artificial Analysis index. One user reported it was surprisingly usable with Codex on an RTX 5090, but Ethan Mollick cautioned that it falls well short of the leading models on agentic tasks and urged people to benchmark it themselves.
- Squeezing it onto smaller machines: Unsloth released new GGUFs with 1-bit quants it says keep about 77% of full accuracy and run on 8GB of RAM. Separately, DFlash 2 reported 70 tokens per second on an M5 Max MacBook Pro, up to 4.6 times faster than standard decoding.
- Open weights cut both ways: Gemma passed 1 billion downloads, a sign of how widely open models are being used. Meanwhile, a refusal-removed Qwen3.8-27B MLX build appeared for Apple Silicon, and even its creators warn it can produce malware, fraud and weapons instructions.
This summary was generated by Claude Sonnet 5.5 to help you explore the links below. Rest assured, I select, organize, and check the links by hand in Google Sheets, and write the introduction and personal commentary in The Main Newsletters myself each week as a labor of love.
This week's links related to Locally Run
Gemma officially achieved 1 billion downloads 🔥 From running underwater to running at space, I’m so excited by how the community is using the models for so many inspiring use cases.”
https://x.com/osanseviero/status/2090490264112738579
Qwen 27B is really good local model but, when you use it, it is immediately absolutely and obviously nowhere near as good as the other models listed here for agentic tasks, and especially for the kinds of complex tasks that GDPval-AA proports to measure Do your own benchmarking!”
https://x.com/emollick/status/2089914722464223706
Qwen3.8-27b is surprisingly usable with codex on my rtx5090! ~200k context with nvfp4 – *very* steerable with skills and plugins – quite smart! great for frontend, basic debugging, and image I don’t trust it (yet) to run in yolo mode – but *very* impressed by how capable it is”
https://x.com/rishdotblog/status/2089458516092399889
512GB RAM Studios. Apple was good to us. 🦞”
https://x.com/steipete/status/2089877190422974974
New open-source agent harness just landed! I got early access to TrueForge by TrueFoundry and have been running it locally for the past few days. The harness layer deserves as much attention as the model, and open source matters here because you can inspect the loop, run it on”
https://x.com/omarsar0/status/2090138030296219973
A “refusal-removed” version of Qwen3.8-27B can now run locally on Apple Silicon. Even its creators warn that it can provide malware, fraud and weapons instructions on demand. It was released as an MLX build in 2, 4, 6 and 8-bit versions. The uploader claims its 4/6/8-bit tests”
https://x.com/kimmonismus/status/2089763435865088508
Ollama has the best performance for deepseek v4 flash on average. For local only, you can try qwen3.8 that is optimized: Apple Silicon: ollama run qwen3.8:27b-mlx NVIDIA: ollama run qwen3.8:27b”
https://x.com/ollama/status/2089454609765146744
Qwen3.8 vs Qwen3.6 vs Gemma 4: 24GB GPU Test
https://kingy.ai/blog/qwen3-8-27b-vs-qwen3-6-27b-vs-gemma-4-31b/
Artificial Analysis Intelligence Index puts Qwen3.8-27B at DeepSeek V4-Pro and GPT 5.6 Luna performance. This is the first time a local model has scored frontier model capability. We weren’t expecting this pace of local progress anywhere near this soon.”
https://x.com/cline/status/2089425906569977896
Qwen3.8-27B is now the #1 local model in Cline after just 4 days. This ends a *4 month* streak by the previous winner Qwen2.5-Coder-7B which has been the top local model since April.”
https://x.com/cline/status/2089825294677143973
We’re releasing new Qwen3.8-27B GGUFs with 10% higher accuracy. Unsloth Dynamic V3 outperforms others by >10% on Div-300, KLD & more benchmarks. We also release 1-bit quants that retain 77% accuracy. Run on 8GB RAM. Blog:
https://t.co/tHsBexyh2K GGUF:
https://x.com/UnslothAI/status/2090103470015828184
Qwen 27B is the “DeepSeek moment” for open source. It matches the closed-source state-of-the-art from just a few months ago and runs on an RTX 5090. Without exaggeration, it’s a game changer.”
https://x.com/kimmonismus/status/2089740575830409700
orcarouter/Qwen3.8-27B-Uncensored-MLX · Hugging Face
https://huggingface.co/orcarouter/Qwen3.8-27B-Uncensored-MLX
We also released 1-bit Qwen3.8-27B quants which run on 8GB RAM. They retain ~77% of accuracy compared to BF16. We originally didn’t want to release them but were shocked that they worked extremely well on our internal testing!”
https://x.com/danielhanchen/status/2090119165055324518
DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro. ⚡ Up to 4.6× the speed of autoregressive decoding, with the same output. This is the next generation of DFlash, seeded at Z Lab and upgraded at Inco AI. Get one more accepted token on every pass, for free!”
https://x.com/zhijianliu_/status/2089836737132650504
EPO: Boosting 3D Foundation Models with Edge-based Pose Optimization” TL;DR: a trackless, differentiable edge-based optimization refines 3D foundation model reconstructions to Bundle Adjustment-level geometry while cutting runtime by up to 80%.”
https://x.com/Almorgand/status/2089746288057204941
LocusGS: Spatially Grounded Tokens for Feed-Forward 3D Gaussian Splatting” TL;DR: uses spatially grounded tokens to organize and decode local Gaussian groups, providing a structured alternative to pixel-aligned feed-forward 3D Gaussian Splatting.”
https://x.com/Almorgand/status/2089376743853380064
Awesome Gemma is live on @github 🤗 A list of awesome Gemma resources, tools, and projects. 1. Model cards and collections for 16 Gemma variants 2. Setup guides for Ollama, vLLM, and LiteRT 3. Fine-tuning recipes for Unsloth, Tunix, and MLX 4. Community tutorials, apps, and”
https://x.com/_philschmid/status/2090485095396180034
When we first introduced @GoogleGemma, our family of open models, our goal was to give developers the tools to build responsible, innovative AI applications anywhere. Today, Gemma models have surpassed a billion downloads. 🎉 Over the past two years, developers have published”
https://x.com/Google/status/2090497445826322464
Kimi K3 is now rolled out to over half of the subscription base for included usage, and we’re continuing to expand access today. US and Europe-hosted and zero data retention.”
https://x.com/ollama/status/2090505028998140182
🐦Ornith 1.5 series models are now available on ollama! 🫡Thanks to our awesome friends at @ollama 👀Come give Ornith 1.5 a try!
https://x.com/ornith_/status/2090276420983587087
Unsloth Dynamic 3.0 GGUFs | Unsloth Documentation
https://unsloth.ai/docs/basics/dynamic-3.0-ggufs
Welcome @ATT to open models!”
https://x.com/ollama/status/2090601698402447748
[2608.16157] FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution
https://arxiv.org/abs/2608.16157
Leave a Reply