Top all-time · #472
Unsloth
Company blog · unsloth.ai/blogaiprogramming
Unsloth is the company blog at unsloth.ai, writing about ai and programming. Since 2023, 121 different Hacker News readers have linked it 218 times, most often “Run DeepSeek R1 Dynamic 1.58-bit”. No feed was found, so read it on the site itself.
Unsloth is an open-source, no-code deskop UI for training and running open models on your local hardware.
Most-linked posts?
- Run DeepSeek R1 Dynamic 1.58-bit13 people · 17 mentions · first linked 2025-01-27
- Qwen3.5 GGUF Benchmarks8 people · 12 mentions · first linked 2026-02-28
- Dynamic 4bit Quantization7 people · 9 mentions · first linked 2024-12-04
- How to run Qwen 3.5 locally7 people · 7 mentions · first linked 2026-02-16
- Qwen3-Coder-Next: How to Run Locally6 people · 7 mentions · first linked 2026-02-03
- /docs/models/qwen3.65 people · 6 mentions · first linked 2026-04-22
- Unsloth GLM-5.2 – How to Run Locally5 people · 5 mentions · first linked 2026-06-19
- Unsloth: Easily finetune and train LLMs Get faster with unsloth5 people · 5 mentions · first linked 2023-12-02
- Unsloth Dynamic 2.0 GGUFs4 people · 7 mentions · first linked 2026-02-04
- Fixing Gemma Bugs4 people · 6 mentions · first linked 2024-03-07
What Hacker News says?
The thing with 3.8 next is that it uses a variant of ngrams. Part of the network is replaced by a lookup table you can store on a fast ssd. In practice, you will be able to run models a bit bigger than 35B. https://unsloth.ai/docs/models/qwen3.8-next
Comment linking /docs/models/qwen3.8-next
on “Mistral Large 4”Submitted Run DeepSeek R1 Dynamic 1.58-bit as a story · 767 points
HN discussionThey were given day zero access to release their quantizations[0]. The lowest quant will still require 256GB VRAM to run inferences, but as always I'm amazed by the Unsloth team by how much they can reduce the requirements for these powerful LLM's, and the easy-to-follow guides to get started with using them. [0] https://unsloth.ai/docs/models/mistral-large-4
Comment linking /docs/models/mistral-large-4
on “We're releasing the weights of Mistral Large 4”Submitted Fine-tune Google's Gemma 3 as a story · 226 points
HN discussionhttps://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B... 816B model, A42B active parameters Unsloth and Meta have guides on how to run the model, but allow me to summarize: You & Me won't be running this one. https://unsloth.ai/docs/models/muse-spark-open
Comment linking /docs/models/muse-spark-open
on “Muse Spark Open”Submitted Phi-4 Bug Fixes as a story · 193 points
HN discussionVision models do degrade more with quantization. https://unsloth.ai/blog/dynamic-4bit
Comment linking /blog/dynamic-4bit
on “Qwen3: Think deeper, act faster”Submitted Fixing Gemma Bugs as a story · 166 points
HN discussionI have similar set-up - can you help out with running it? Was it in ollama? EDIT: It seems that original authors provided a nice write-up: https://unsloth.ai/blog/deepseekr1-dynamic#:~:text=%F0%9F%96...
Comment linking /blog/deepseekr1-dynamic
on “Run DeepSeek R1 Dynamic 1.58-bit”Submitted Bugs in LLM Training – Gradient Accumulation Fix as a story · 81 points
HN discussionOh I wrote a bit about it in https://docs.unsloth.ai/basics/unsloth-dynamic-2.0-ggufs and https://unsloth.ai/blog/deepseekr1-dynamic if that helps!
Comment linking /blog/deepseekr1-dynamic
on “Qwen3-Coder: Agentic coding in the world”Submitted Long-Context GRPO as a story · 60 points
HN discussionOh we actually fixed bugs! We fixed a few bugs in Gemma - see https://news.ycombinator.com/item?id=39671146 , a gradient accumulation bug see https://news.ycombinator.com/item?id=41859037 , Phi bugs, Llama bugs and more! See https://unsloth.ai/blog/reintroducing for more details!
Comment linking /blog/reintroducing
on “Qwen3-Coder: Agentic coding in the world”Submitted Run DeepSeek R1 Dynamic 1.58-bit as a story · 19 points
HN discussionI wrote approximately in the blog about it and linked some papers! I also wrote about it here - https://unsloth.ai/blog/dynamic-4bit - one has to inspect the activation and weight quantization errors!
Comment linking /blog/dynamic-4bit
on “Qwen3-Coder: Agentic coding in the world”Submitted Train your own R1 reasoning model as a story · 11 points
HN discussionThere were initial difficulties in finetuning that made it less appealing early on, and that's snowballed a bit into having more of a focus on RAG. Some of the issues still exist, of course: * Finetuning takes time and compute; for one-off queries using in-context learning is vastly more efficient (i.e., look it up with RAG). * Early results with finetuning had trouble reliably memorizing information. We've got a mu…
Comment linking /blog/contpretraining
on “Claude Integrations”Submitted Train your own R1 reasoning model with Unsloth as a story · 6 points
HN discussionEarlier today I read a reddit comment[1] about a guy who tried running the quantized version from unsloth[2] on 4xH100 and the results was underwhelming (it ended up costing $137 per 1 million tokens). Any idea of what they're doing wrong? [1]: https://www.reddit.com/r/LocalLLaMA/comments/1icphqa/how_to_... [2]: https://unsloth.ai/blog/deepseekr1-dynamic
Comment linking /blog/deepseekr1-dynamic
on “An analysis of DeepSeek's R1-Zero and R1”Submitted Qwen3.8-Flash-Next: How to Run Locally as a story · 5 points
HN discussionWhy did you use 1-bit quantization vs 3-bit quantization? It looks like the 3-bit requires 90 GB[1] which, I imagine, would fit within the DGX Spark's 128GB of unified memory. [1] https://unsloth.ai/docs/models/qwen3.8-next
Comment linking /docs/models/qwen3.8-next
on “Qwen3.8-Flash-Next”Submitted Claude Code makes local LLMs 90% slower as a story · 4 points
HN discussionIn their page they say it will need at least 112GB[0], so including context, that would be a tight fit. I'm also hoping I can make a q4 fit on my 128GB strix halo [0]: https://unsloth.ai/docs/models/qwen3.8-next#qwen3.8-flash-ne...
Comment linking /docs/models/qwen3.8-next
on “Qwen3.8-Flash-Next”Submitted Long context GPT-OSS fine-tuning as a story · 4 points
HN discussionThey are already working on it. https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF https://unsloth.ai/docs/models/qwen3.8-next > You will need at least 75 GB of RAM or unified memory to run the model. Its smallest 1-bit quantized version is larger than usual because of the model’s architecture so 1-bit isn't really 1-bit at all. However, this also means the quantization is less aggressive, allowing the model to…
Comment linking /docs/models/qwen3.8-next
on “Qwen3.8-Flash-Next”Submitted Unsloth creators fix universal error with gradient accumulation as a story · 4 points
HN discussionNice! Does anyone know how this compares to the Unsloth quantizations of this model? https://unsloth.ai/docs/models/qwen3.8#run-qwen3.8-guide
Comment linking /docs/models/qwen3.8
on “Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint”Submitted Unsloth - Qwen3.8 - How to Run Locally as a story · 3 points
HN discussionAlternate idea: Download Unsloth Studio https://unsloth.ai/docs/new/studio#quickstart From the command line: unsloth studio Browse to the URL it gives you (eg http://127.0.0.1:8888/ ) Download a good fast local model e.g. Gemma-4-12B (small machine) or Qwen3.6-35B-A3B (larger memory). Or a Nemotron model, if you prefer that. Then you'll have both your backup/emergency assistant that runs on your own computer and be…
Comment linking /docs/new/studio
on “Survival LLM: AI assistant and emergency library that runs on your own computer”Submitted Unsloth GLM-5.2 – How to Run Locally as a story · 3 points
HN discussion
