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.
This discussion details performance benchmarks of llama.cpp on an NVIDIA DGX Spark, including tests for various models (gpt-oss-20b, gpt-oss-120b, Qwen3, Qwen2.5, Gemma, GLM) with different context depths and batch sizes.
Extracting structured information effectively and accurately from long unstructured text with LangExtract and LLMs. This article explores Google’s LangExtract framework and its open-source LLM, Gemma 3, demonstrating how to parse an insurance policy to surface details like exclusions.
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.
This article details the release of Gemma 3, the latest iteration of Google’s open-weights language model. Key improvements include **vision-language capabilities** (using a tailored SigLIP encoder), **increased context length** (up to 128k tokens for larger models), and **architectural changes for improved memory efficiency** (5-to-1 interleaved attention and removal of softcapping). Gemma 3 demonstrates superior performance compared to Gemma 2 across benchmarks and offers models optimized for various use cases, including on-device applications with the 1B model.
A Reddit thread discussing preferred local Large Language Model (LLM) setups for tasks like summarizing text, coding, and general use. Users share their model choices (Gemma, Qwen, Phi, etc.) and frameworks (llama.cpp, Ollama, EXUI) along with potential issues and configurations.
| **Model** | **Use Cases** | **Size (Parameters)** | **Approx. VRAM (Q4 Quantization)** | **Approx. RAM (Q4)** | **Notes/Requirements** |
|----------------|---------------------------------------------------|------------------------|-----------------------------------|---------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| **Gemma 3 (Meta)** | Summarization, conversational tasks, image recognition, translation, simple writing | 3B, 4B, 7B, 8B, 12B, 27B+ | 2-4GB (3B), 4-6GB (7B), 8-12GB (12B) | 4-8GB (3B), 8-12GB (7B), 16-24GB (12B) | Excellent performance for its size. Recent versions have had memory leak issues (see Reddit post – use Ollama 0.6.6 or later, but even that may not be fully fixed). QAT versions are highly recommended. |
| **Qwen 2.5 (Alibaba)** | Summarization, coding, reasoning, decision-making, technical material processing | 3.5B, 7B, 72B | 2-3GB (3.5B), 4-6GB (7B), 26-30GB (72B) | 4-6GB (3.5B), 8-12GB (7B), 50-60GB (72B) | Qwen models are known for strong performance. Coder versions specifically tuned for code generation. |
| **Qwen3 (Alibaba - upcoming)**| General purpose, likely similar to Qwen 2.5 with improvements | 70B | Estimated 25-30GB (Q4) | 50-60GB | Expected to be a strong competitor. |
| **Llama 3 (Meta)**| General purpose, conversation, writing, coding, reasoning | 8B, 13B, 70B+ | 4-6GB (8B), 7-9GB (13B), 25-30GB (70B) | 8-12GB (8B), 14-18GB (13B), 50-60GB (70B) | Current state-of-the-art open-source model. Excellent balance of performance and size. |
| **YiXin (01.AI)** | Reasoning, brainstorming | 72B | ~26-30GB (Q4) | ~50-60GB | A powerful model focused on reasoning and understanding. Similar VRAM requirements to Qwen 72B. |
| **Phi-4 (Microsoft)** | General purpose, writing, coding | 14B | ~7-9GB (Q4) | 14-18GB | Smaller model, good for resource-constrained environments, but may not match larger models in complexity. |
| **Ling-Lite** | RAG (Retrieval-Augmented Generation), fast processing, text extraction | Variable | Varies with size | Varies with size | MoE (Mixture of Experts) model known for speed. Good for RAG applications where quick responses are important. |
**Key Considerations:**
* **Quantization:** The VRAM and RAM estimates above are based on 4-bit quantization (Q4). Lower quantization (e.g., Q2) will reduce memory usage further, but *may* impact quality. Higher quantization (e.g., Q8, FP16) will increase quality but require significantly more memory.
* **Frameworks:** Popular frameworks for running these models locally include:
* **llama.cpp:** Highly optimized for CPU and GPU, especially on Apple Silicon.
* **Ollama:** Simplified setup and management of LLMs. (Be aware of the Gemma 3 memory leak issue!)
* **Text Generation WebUI (oobabooga):** Web-based interface with many features and customization options.
* **Hardware:** A dedicated GPU with sufficient VRAM is highly recommended for decent performance. CPU-only inference is possible but can be slow. More RAM is generally better, even if the model fits in VRAM.
* **Context Length:** The "40k" context mentioned in the Reddit post refers to the maximum number of tokens (words or sub-words) the model can process at once. Longer context lengths require more memory.
| **Model** | **Use Cases** | **Size (Parameters)** | **Approx. VRAM (Q4 Quantization)** | **Approx. RAM (Q4)** | **Notes/Requirements** |
|----------------|---------------------------------------------------|------------------------|-----------------------------------|---------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| **Gemma 3 (Meta)** | Summarization, conversational tasks, image recognition, translation, simple writing | 3B, 4B, 7B, 8B, 12B, 27B+ | 2-4GB (3B), 4-6GB (7B), 8-12GB (12B) | 4-8GB (3B), 8-12GB (7B), 16-24GB (12B) | Excellent performance for its size. Recent versions have had memory leak issues (see Reddit post – use Ollama 0.6.6 or later, but even that may not be fully fixed). QAT versions are highly recommended. |
| **Qwen 2.5 (Alibaba)** | Summarization, coding, reasoning, decision-making, technical material processing | 3.5B, 7B, 72B | 2-3GB (3.5B), 4-6GB (7B), 26-30GB (72B) | 4-6GB (3.5B), 8-12GB (7B), 50-60GB (72B) | Qwen models are known for strong performance. Coder versions specifically tuned for code generation. |
| **Qwen3 (Alibaba - upcoming)**| General purpose, likely similar to Qwen 2.5 with improvements | 70B | Estimated 25-30GB (Q4) | 50-60GB | Expected to be a strong competitor. |
| **Llama 3 (Meta)**| General purpose, conversation, writing, coding, reasoning | 8B, 13B, 70B+ | 4-6GB (8B), 7-9GB (13B), 25-30GB (70B) | 8-12GB (8B), 14-18GB (13B), 50-60GB (70B) | Current state-of-the-art open-source model. Excellent balance of performance and size. |
| **YiXin (01.AI)** | Reasoning, brainstorming | 72B | ~26-30GB (Q4) | ~50-60GB | A powerful model focused on reasoning and understanding. Similar VRAM requirements to Qwen 72B. |
| **Phi-4 (Microsoft)** | General purpose, writing, coding | 14B | ~7-9GB (Q4) | 14-18GB | Smaller model, good for resource-constrained environments, but may not match larger models in complexity. |
| **Ling-Lite** | RAG (Retrieval-Augmented Generation), fast processing, text extraction | Variable | Varies with size | Varies with size | MoE (Mixture of Experts) model known for speed. Good for RAG applications where quick responses are important. |
**Key Considerations:**
* **Quantization:** The VRAM and RAM estimates above are based on 4-bit quantization (Q4). Lower quantization (e.g., Q2) will reduce memory usage further, but *may* impact quality. Higher quantization (e.g., Q8, FP16) will increase quality but require significantly more memory.
* **Frameworks:** Popular frameworks for running these models locally include:
* **llama.cpp:** Highly optimized for CPU and GPU, especially on Apple Silicon.
* **Ollama:** Simplified setup and management of LLMs. (Be aware of the Gemma 3 memory leak issue!)
* **Text Generation WebUI (oobabooga):** Web-based interface with many features and customization options.
* **Hardware:** A dedicated GPU with sufficient VRAM is highly recommended for decent performance. CPU-only inference is possible but can be slow. More RAM is generally better, even if the model fits in VRAM.
* **Context Length:** The "40k" context mentioned in the Reddit post refers to the maximum number of tokens (words or sub-words) the model can process at once. Longer context lengths require more memory.
This document details how to run Gemma models, covering framework selection, variant choice, and running generation/inference requests. It emphasizes considering available hardware resources and provides recommendations for beginners.
This article details a method for converting PDFs to Markdown using a local LLM (Gemma 3 via Ollama), focusing on privacy and efficiency. It involves rendering PDF pages as images and then using the LLM for content extraction, even from scanned PDFs.
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.
Gemma Scope is an open-source, multi-scale, high-throughput microscope system that combines brightfield, fluorescence, and confocal microscopy, designed for imaging large samples like brain tissue.
