This article details research into finding the optimal architecture for small language models (70M parameters), exploring depth-width tradeoffs, comparing different architectures, and introducing Dhara-70M, a diffusion model offering 3.8x faster throughput with improved factuality.
A Vim plugin that provides local LLM-assisted code and text completion using llama.cpp server instances. It supports features like auto-suggest on cursor movement, manual toggle with Ctrl+F, accepting suggestions with Tab/Shift+Tab, configurable context scope, and performance stats display.
This blog post details the training of 'Chess Llama', a small Llama model designed to play chess. It covers the inspiration behind the project (Chess GPT), the dataset used (Lichess Elite database), the training process using Huggingface Transformers, and the model's performance (Elo rating of 1350-1400). It also includes links to try the model and view the source code.
A detailed comparison of the architectures of recent large language models (LLMs) including DeepSeek-V3, OLMo 2, Gemma 3, Mistral Small 3.1, Llama 4, Qwen3, SmolLM3, and Kimi 2, focusing on key design choices and their impact on performance and efficiency.
1. **DeepSeek V3/R1**:
- Uses Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE) for efficiency.
- MLA compresses key and value tensors to reduce KV cache memory usage.
- MoE activates only a subset of experts per token, improving inference efficiency.
2. **OLMo 2**:
- Focuses on transparency in training data and code.
- Uses RMSNorm layers placed after attention and feed-forward modules (Post-Norm).
- Introduces QK-Norm, an additional RMSNorm layer applied to queries and keys inside the attention mechanism.
3. **Gemma 3**:
- Employs sliding window attention to reduce memory requirements in the KV cache.
- Uses a 5:1 ratio of sliding window attention to global attention layers.
- Combines Pre-Norm and Post-Norm RMSNorm layers around the attention module.
4. **Mistral Small 3.1**:
- Outperforms Gemma 3 27B on several benchmarks while being faster.
- Uses a standard architecture with a custom tokenizer and reduced KV cache and layer count.
5. **Llama 4**:
- Adopts an MoE approach similar to DeepSeek V3 but with fewer, larger experts.
- Alternates MoE and dense modules in every other transformer block.
6. **Qwen3**:
- Comes in both dense and MoE variants.
- Dense models are easier to fine-tune and deploy, while MoE models are optimized for scaling inference.
7. **SmolLM3**:
- Uses No Positional Embeddings (NoPE), omitting explicit positional information injection.
- NoPE improves length generalization, meaning performance deteriorates less with increased sequence length.
8. **Kimi K2 and Kimi K2 Thinking**:
- Uses a variant of the Muon optimizer over AdamW.
- Kimi K2 Thinking extends the context size to 256k tokens.
9. **GPT-OSS**:
- OpenAI's first open-weight models since GPT-2.
- Uses sliding window attention and a width-versus-depth trade-off.
10. **Grok 2.5**:
- Uses a small number of large experts and a shared expert module.
- Reflects an older trend in MoE architectures.
11. **GLM-4.5**:
- Comes in two variants: a 355-billion-parameter model and a more compact 106-billion-parameter version.
- Uses a shared expert and starts with several dense layers before introducing MoE blocks.
12. **Qwen3-Next**:
- Introduces a Gated DeltaNet + Gated Attention hybrid mechanism.
- Uses Multi-Token Prediction (MTP) for efficiency.
13. **MiniMax-M2**:
- Uses per-layer QK-Norm and partial RoPE.
- More "sparse" than Qwen3, with fewer active experts per token.
14. **Kimi Linear**:
- Modifies the linear attention mechanism with Kimi Delta Attention (KDA).
- Combines Gated DeltaNet with Multi-Head Latent Attention (MLA).
15. **Olmo 3 Thinking**:
- Uses sliding window attention and YaRN for context extension.
- Comes in base, instruct, and reasoning variants.
16. **DeepSeek V3.2**:
- Adds a sparse attention mechanism to improve efficiency.
- On par with GPT-5.1 and Gemini 3.0 Pro on certain benchmarks.
17. **Mistral 3**:
- First MoE model since Mixtral in 2023.
- Partnered with NVIDIA for optimization on Blackwell chips.
18. **Nemotron 3**:
- A Transformer-Mamba hybrid architecture.
- Interleaves Mamba-2 sequence-modeling blocks with sparse MoE feed-forward layers.
19. **Xiaomi MiMo-V2-Flash**:
- Uses sliding window attention in a 5:1 ratio with global attention.
- Employs multi-token prediction (MTP) for efficiency.
20. **Arcee AI Trinity Large**:
- Uses alternating local:global attention layers, NoPE, and gated attention.
- Introduces depth-scaled sandwich norm for training stability.
1. **DeepSeek V3/R1**:
- Uses Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE) for efficiency.
- MLA compresses key and value tensors to reduce KV cache memory usage.
- MoE activates only a subset of experts per token, improving inference efficiency.
2. **OLMo 2**:
- Focuses on transparency in training data and code.
- Uses RMSNorm layers placed after attention and feed-forward modules (Post-Norm).
- Introduces QK-Norm, an additional RMSNorm layer applied to queries and keys inside the attention mechanism.
3. **Gemma 3**:
- Employs sliding window attention to reduce memory requirements in the KV cache.
- Uses a 5:1 ratio of sliding window attention to global attention layers.
- Combines Pre-Norm and Post-Norm RMSNorm layers around the attention module.
4. **Mistral Small 3.1**:
- Outperforms Gemma 3 27B on several benchmarks while being faster.
- Uses a standard architecture with a custom tokenizer and reduced KV cache and layer count.
5. **Llama 4**:
- Adopts an MoE approach similar to DeepSeek V3 but with fewer, larger experts.
- Alternates MoE and dense modules in every other transformer block.
6. **Qwen3**:
- Comes in both dense and MoE variants.
- Dense models are easier to fine-tune and deploy, while MoE models are optimized for scaling inference.
7. **SmolLM3**:
- Uses No Positional Embeddings (NoPE), omitting explicit positional information injection.
- NoPE improves length generalization, meaning performance deteriorates less with increased sequence length.
8. **Kimi K2 and Kimi K2 Thinking**:
- Uses a variant of the Muon optimizer over AdamW.
- Kimi K2 Thinking extends the context size to 256k tokens.
9. **GPT-OSS**:
- OpenAI's first open-weight models since GPT-2.
- Uses sliding window attention and a width-versus-depth trade-off.
10. **Grok 2.5**:
- Uses a small number of large experts and a shared expert module.
- Reflects an older trend in MoE architectures.
11. **GLM-4.5**:
- Comes in two variants: a 355-billion-parameter model and a more compact 106-billion-parameter version.
- Uses a shared expert and starts with several dense layers before introducing MoE blocks.
12. **Qwen3-Next**:
- Introduces a Gated DeltaNet + Gated Attention hybrid mechanism.
- Uses Multi-Token Prediction (MTP) for efficiency.
13. **MiniMax-M2**:
- Uses per-layer QK-Norm and partial RoPE.
- More "sparse" than Qwen3, with fewer active experts per token.
14. **Kimi Linear**:
- Modifies the linear attention mechanism with Kimi Delta Attention (KDA).
- Combines Gated DeltaNet with Multi-Head Latent Attention (MLA).
15. **Olmo 3 Thinking**:
- Uses sliding window attention and YaRN for context extension.
- Comes in base, instruct, and reasoning variants.
16. **DeepSeek V3.2**:
- Adds a sparse attention mechanism to improve efficiency.
- On par with GPT-5.1 and Gemini 3.0 Pro on certain benchmarks.
17. **Mistral 3**:
- First MoE model since Mixtral in 2023.
- Partnered with NVIDIA for optimization on Blackwell chips.
18. **Nemotron 3**:
- A Transformer-Mamba hybrid architecture.
- Interleaves Mamba-2 sequence-modeling blocks with sparse MoE feed-forward layers.
19. **Xiaomi MiMo-V2-Flash**:
- Uses sliding window attention in a 5:1 ratio with global attention.
- Employs multi-token prediction (MTP) for efficiency.
20. **Arcee AI Trinity Large**:
- Uses alternating local:global attention layers, NoPE, and gated attention.
- Introduces depth-scaled sandwich norm for training stability.
Transformer Lab is an open-source application for advanced LLM engineering, allowing users to interact, train, fine-tune, and evaluate large language models on their own computer. It supports various models, hardware, and inference engines and includes features like RAG, dataset building, and a REST API.
Simon Willison explains how to use the mistral.rs library in Rust to run the Llama Vision model on a Mac M2 laptop. He provides a detailed example and discusses the memory usage and GPU utilization.
This article compares the performance of smaller language models Gemma, Llama 3, and Mistral on reading comprehension tasks. The author highlights the trend of smaller, more accessible models and discusses Apple's recent foray into the field with its own proprietary model.
This blog post benchmarks and compares the performance of SGLang, TensorRT-LLM, and vLLM for serving large language models (LLMs). SGLang demonstrates superior or competitive performance in offline and online scenarios, often outperforming vLLM and matching or exceeding TensorRT-LLM.
A discussion post on Reddit's LocalLLaMA subreddit about logging the output of running models and monitoring performance, specifically for debugging errors, warnings, and performance analysis. The post also mentions the need for flags to output logs as flat files, GPU metrics (GPU utilization, RAM usage, TensorCore usage, etc.) for troubleshooting and analytics.
This article provides a beginner-friendly introduction to Large Language Models (LLMs) and explains the key concepts in a clear and organized way.
