BitShovel Workbench · Open-source projects https://bitshovel.site/en/ Growth signals from open models and tooling: releases, trend charts, ecosystem expansion. en-US Sun, 06 Sep 2026 10:03:38 UTC MiniMind: learn how a small language model is trained https://bitshovel.site/en/card/github-weekly-minimind https://bitshovel.site/en/card/github-weekly-minimind Sun, 06 Sep 2026 01:11:43 UTC MiniMind provides a roughly 64M-parameter model and training tutorials. Its “two hours” refers to one supervised fine-tuning epoch on a single RTX 3090. MiniMind provides a roughly 64M-parameter model and training tutorials. Its “two hours” refers to one supervised fine-tuning epoch on a single RTX 3090.

Original: https://github.com/jingyaogong/minimind

]]> Qwen3.8-27B is entry #2 in this filtered HF trend collection https://bitshovel.site/en/card/hf-trending-qwen-qwen3-8-27b-92bb21fbd5739d6d https://bitshovel.site/en/card/hf-trending-qwen-qwen3-8-27b-92bb21fbd5739d6d Sat, 05 Sep 2026 10:01:59 UTC In the 2026-09-05 HF trend collection, Qwen/Qwen3.8-27B is entry #2 after task, access, and score filters. This is not its original HF chart rank. The snapshot reports a trending score of 551, 6,024,467 downloads, and 13,984 likes. In the 2026-09-05 HF trend collection, Qwen/Qwen3.8-27B is entry #2 after task, access, and score filters. This is not its original HF chart rank. The snapshot reports a trending score of 551, 6,024,467 downloads, and 13,984 likes.

Original: https://huggingface.co/Qwen/Qwen3.8-27B

]]> Help coding agents write Go for your project version https://bitshovel.site/en/card/github-weekly-go-modern-guidelines https://bitshovel.site/en/card/github-weekly-go-modern-guidelines Thu, 03 Sep 2026 22:00:28 UTC JetBrains’ guidelines tell coding agents to read go.mod before choosing language and standard-library features. They aim to avoid outdated patterns and code that requires a newer Go version than the project uses. JetBrains’ guidelines tell coding agents to read go.mod before choosing language and standard-library features. They aim to avoid outdated patterns and code that requires a newer Go version than the project uses.

Original: https://github.com/JetBrains/go-modern-guidelines

]]> funes helps coding agents recall past decisions https://bitshovel.site/en/card/feed-story-huggingface-co-blog-funes https://bitshovel.site/en/card/feed-story-huggingface-co-blog-funes Thu, 03 Sep 2026 16:20:33 UTC funes indexes sessions from Claude Code, Codex and other agents locally, with results linked to the original conversation. Optional publishing to a Hugging Face dataset supports access across machines or collaborators. funes indexes sessions from Claude Code, Codex and other agents locally, with results linked to the original conversation. Optional publishing to a Hugging Face dataset supports access across machines or collaborators.

Original: https://huggingface.co/blog/funes

]]> Training a coding model to paint watercolours with TRL and OpenEnv https://bitshovel.site/en/card/feed-story-huggingface-co-blog-train-to-paint-with-code https://bitshovel.site/en/card/feed-story-huggingface-co-blog-train-to-paint-with-code Thu, 03 Sep 2026 10:00:23 UTC Hugging Face blog published "Training a coding model to paint watercolours with TRL and OpenEnv". The line above and the linked original carry the substance; judge conclusions on the original's own evidence. Hugging Face blog published "Training a coding model to paint watercolours with TRL and OpenEnv". The line above and the linked original carry the substance; judge conclusions on the original's own evidence.

Original: https://huggingface.co/blog/train-to-paint-with-code

]]> Lily: a Mac inference server for one model https://bitshovel.site/en/card/alphasignal-alphasignal-ai-news-perplexity-s-lily-beats-mlx https://bitshovel.site/en/card/alphasignal-alphasignal-ai-news-perplexity-s-lily-beats-mlx Thu, 03 Sep 2026 05:50:49 UTC Perplexity’s open-source Lily is a narrow implementation for studying local inference. It accepts a specific Qwen3.6-35B-A3B 4-bit checkpoint, requires M5-or-newer hardware and macOS 26 or later, and must be built from source. Its text-only API does not support streaming, tool calls or multimodal inputs. Perplexity’s open-source Lily is a narrow implementation for studying local inference. It accepts a specific Qwen3.6-35B-A3B 4-bit checkpoint, requires M5-or-newer hardware and macOS 26 or later, and must be built from source. Its text-only API does not support streaming, tool calls or multimodal inputs.

Original: https://github.com/perplexityai/pplx-garden/tree/main/lily

]]> Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI https://bitshovel.site/en/card/feed-story-huggingface-co-blog-webgpu-kernels https://bitshovel.site/en/card/feed-story-huggingface-co-blog-webgpu-kernels Wed, 02 Sep 2026 05:42:07 UTC Hugging Face's blog announces @huggingface/kernels: 200+ WebGPU kernels for local AI in the browser, callable from a web page without a server (AIHOT's listing counts 207). Open-source, official release, with the full kernel list and usage in the post. Hugging Face's blog announces @huggingface/kernels: 200+ WebGPU kernels for local AI in the browser, callable from a web page without a server (AIHOT's listing counts 207). Open-source, official release, with the full kernel list and usage in the post.

Original: https://huggingface.co/blog/webgpu-kernels

]]> OpenMAIC turns materials into editable interactive lessons https://bitshovel.site/en/card/github-weekly-openmaic https://bitshovel.site/en/card/github-weekly-openmaic Thu, 27 Aug 2026 13:13:53 UTC OpenMAIC generates slides, quizzes and interactive activities from a topic or source material. Version 1.0, released August 27, adds conversational course editing and resumable sessions when server storage is configured. OpenMAIC generates slides, quizzes and interactive activities from a topic or source material. Version 1.0, released August 27, adds conversational course editing and resumable sessions when server storage is configured.

Original: https://github.com/THU-MAIC/OpenMAIC

]]> DeepSeek releases experimental V4-Flash-Vision-Exp https://bitshovel.site/en/card/deepseek-v4-flash-vision-exp https://bitshovel.site/en/card/deepseek-v4-flash-vision-exp Tue, 01 Sep 2026 01:32:15 UTC DeepSeek adds visual understanding to V4-Flash and provides weights, prompt-encoding references, and minimal inference code under MIT. The experimental release includes multimodal-agent benchmarks and deployment examples. DeepSeek adds visual understanding to V4-Flash and provides weights, prompt-encoding references, and minimal inference code under MIT. The experimental release includes multimodal-agent benchmarks and deployment examples.

Original: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

]]> Qwen3.8-Flash-Next weights offer a Qwen4 architecture preview https://bitshovel.site/en/card/aihot-aihot-virxact-com-items-cmta2veap03nmrolwxllvp4a https://bitshovel.site/en/card/aihot-aihot-virxact-com-items-cmta2veap03nmrolwxllvp4a Thu, 27 Aug 2026 00:42:11 UTC Qwen provides an experimental multimodal MoE preview and an FP8 variant. Its model card now specifies 125B language-model parameters with 6B active, plus 51B n-gram embeddings and 4B MTP parameters, under Qwen Community License 1.0. Qwen provides an experimental multimodal MoE preview and an FP8 variant. Its model card now specifies 125B language-model parameters with 6B active, plus 51B n-gram embeddings and 4B MTP parameters, under Qwen Community License 1.0.

Original: https://qwen.ai/blog?id=qwen3.8-flash-next

]]>