DeepSeek's V4 Flash 0731, a 284B-parameter Mixture-of-Experts model with 13B active parameters and a native 1M-token context under an MIT license, is cheaper to use via the API than to run locally, ac…
Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter sparse Mixture-of-Experts model, on August 3, claiming it outperforms GPT-5.6 Sol on key coding benchmarks and promising full open weights next w…
Independent researcher Alpamys Makazhan released Soup, a Show HN project that fine-tunes a full Llama-3.1-8B model in NF4 quantization with a 3.32 GB VRAM peak at 119.6 tokens/sec on a 4 GB RTX 3050 L…
Unsloth, an open-source AI startup, released Unsloth Studio (Beta), a platform that lets users run and train text, audio, embedding, and vision models locally on Windows, Linux, and macOS, with suppor…
DeepSeek released V4 Flash 0731, an iterative update of its Mixture-of-Experts model with 284B total parameters and 13B active per token, achieving an Artificial Analysis Intelligence Index of 50, up …
Z.ai's GLM-5.2 scored 62.1 on SWE-bench Pro, beating GPT-5.5's 58.6, while costing $1.40 per million input tokens and $4.40 per million output tokens, making it 3.6x cheaper on input and 5.7x cheaper …
DeepSeek released DeepSeek-V4-Flash-0731 on Hugging Face and moved its V4-Flash API to public beta on July 31, 2026, claiming major gains in agentic and coding benchmarks from re-post-training rather …
The accuracy gap between the best open-weight models and GPT-4o has shrunk to under 3% on structured data tasks, according to a developer's hands-on analysis. A fine-tuned Qwen2.5 72B model achieved 9…
A first-principles analysis of LoRA fine-tuning memory usage reveals that 87.3% of VRAM scaling with sequence length for Llama 3.1 8B is consumed by the cross-entropy loss head tensor, not the model, …
Benchmarking Qwen 3.6 35B MoE (3B active) on an RTX 3090 using Unsloth's UD-IQ4_NL_XL quantisation achieved up to 140 tokens per second for generation and over 3,300 tok/s for prompt processing with a…
Unsloth achieves up to 7.3x training speedup over standard Transformers for MoE models like gpt-oss-20B on an NVIDIA B200, according to published benchmarks, while Axolotl delivers up to 1.45x speedup…
AMD's Ryzen AI Developer Center, bundled with Ryzen AI Max Plus 395 machines, eliminates manual ROCm and driver setup for local AI workloads, offering guided playbooks for ComfyUI, LM Studio, and Unsl…
Unsloth announced AMD GPU support for local LLM training and inference, enabling 500+ models to run up to 2× faster with 70% less VRAM on AMD Radeon, Instinct, Ryzen, and data center GPUs. The release…
Thinking Machines' open Inkling model, with 975 billion total parameters and 41 billion active, fits on a single high-memory box at 2-bit or 3-bit quantization, unlike the datacenter-scale Kimi K3. Re…
A developer deploying a fine-tuned Llama 3.2 3B medical AI model offline on Android for Yoruba, Hausa, Igbo, and Nigerian Pidgin encountered three distinct crashes during GGUF conversion—tokenizer mis…
Researchers fine-tuned LLaMA 3 (8B) as a drop-in reranker for RAG pipelines, achieving gains of 14% in answer relevancy, 16% in context precision, 19% in answer similarity, and 21% in answer correctne…
A student fine-tuned Meta's Llama 3.1 8B model for multi-step mathematical reasoning using Unsloth, LoRA, and a 'Silent Coder' approach, all within the RAM limits of a free Google Colab instance with …
A new course chapter introduces reinforcement learning (RL) and its application to training large language models (LLMs), explaining core concepts such as agent, environment, action, reward, and polic…
A developer detailed the local deployment of GLM 5.2, a 753B MoE model with 40B active parameters requiring 1.51TB of RAM in BF16 precision. Unsloth's dynamic quantizations, such as UD-Q3_K_XL at 343G…
Amazon Web Services announced a partnership with Unsloth to deploy quantized foundation models on Amazon SageMaker AI, reducing memory usage and serving costs while maintaining accuracy. The collabora…