New blocks

Llama-3.2-1B-Instruct is Meta's compact multilingual model at 1B parameters. This instruction-tuned version balances efficiency with generative capabilities across multiple languages.

Qwen-2-VL-72B-Instruct is a multimodal model handling images, videos, and text across languages. This 72B parameter model excels in visual analysis, 20+ minute video understanding, device control, and multilingual text recognition.

L3-8B-Lunaris merges Llama-3-8B variants for general tasks and creative roleplay. This 8B parameter model by Sao10K balances broad capabilities with narrative generation.

Midnight-Rose-70B merges multiple models for advanced creative writing and storytelling. This 70B parameter model builds on Rogue Rose and Aurora Nights, optimizing for detailed narrative generation and character interactions.

OpenHermes-2.5-Mistral-7B enhances the Mistral-7B base model with specialized code training. This iteration builds on OpenHermes 2's foundation while expanding programming capabilities.

Nous-Hermes-Llama2-13B fine-tunes Llama-2-13B on 300K curated instructions. This collaborative effort between Nous Research, Teknium, and Emozilla optimizes instruction-following capabilities.

Airoboros-L2-70B fine-tunes Llama-2-70B for extended creative writing and narrative generation. This model specializes in chapter continuations, character dialogues, and multi-participant storytelling scenarios.

Dolphin-Mixtral-8x22B fine-tunes Mixtral's MoE architecture with 64k context length. This model enhances instruction following, conversation, and coding capabilities while requiring external alignment for deployment.

L3-70B-Euryale-v2.1 is a creative writing model based on Llama-3-70B, optimized for narrative generation and character interactions. This 70B parameter model focuses on imaginative storytelling and dynamic dialogue.

Hermes-2-Pro-Llama-3-8B fine-tunes Llama-3-8B with the enhanced OpenHermes 2.5 dataset. This 8B model specializes in function calling and structured JSON outputs while maintaining strong general capabilities.

OpenChat-7B uses C-RLFT (Conditioned Reinforcement Learning Fine-Tuning) to enhance a 7B parameter base model. This innovative approach trains on mixed-quality data without requiring preference labels.

Llama-3-70B-Instruct is Meta's instruction-tuned model at 70B parameters, trained on 15T tokens. This decoder-only architecture shows competitive dialogue performance against leading closed-source models.

Mistral-7B-Instruct fine-tunes the base Mistral-7B for instruction following. This 7.3B parameter model optimizes speed and context length while maintaining strong performance across diverse tasks.

Gemma-2-9B-IT is Google's instruction-tuned lightweight model based on Gemini technology. This 9B parameter decoder-only model balances performance with efficiency for diverse text-generation tasks.

WizardLM-2 8x22B is Microsoft's latest Mixture-of-Experts model, using 8 expert networks of 22B parameters each. This 176B total parameter architecture delivers competitive performance in knowledge-intensive tasks.

Llama-3-8B-Instruct is Meta's compact dialogue model with 8B parameters, trained on 15T tokens. This instruction-tuned decoder demonstrates competitive performance against larger closed-source models in human evaluations.

Qwen-2.5-72B-Instruct leads the Qwen2.5 series at 72B parameters. This instruction-tuned transformer excels in language understanding, coding, mathematics, reasoning, and multilingual tasks.

MythoMax L2 13B combines MythoLogic-L2's comprehension with Huginn's writing capabilities through innovative tensor merging. This 13B hybrid model balances robust understanding with creative expression.

L3-8B-Stheno-v3.2 is an 8B parameter model optimized for dynamic role-play and character immersion. This specialized model focuses on consistent character portrayal and contextual responses.

DeepSeek R1 Distill Qwen 32B distils Qwen 2.5 and DeepSeek R1 expertise into a powerful 32B model. Outperforms OpenAI's o1-mini with impressive math capabilities (AIME: 72.6%, MATH-500: 94.3%) and coding skills (CodeForces: 1691).

DeepSeek R1 Distill Qwen 14B leverages Qwen 2.5 and DeepSeek R1 for superior performance. This 14B distilled model outperforms OpenAI's o1-mini, scoring 74% on MMLU and excelling in math tasks (AIME: 69.7%, MATH-500: 93.9%)

本地deepseek-r1

Mistral-Nemo combines NVIDIA collaboration with impressive 128k context in a 12B package. This Apache-licensed model supports 11 languages and function calling, while outperforming similarly-sized competitors.

Meta's Llama 3.1 70B-instruct model excels in multilingual dialogue and complex reasoning. This instruction-tuned powerhouse sits between the efficient 8B and massive 405B variants, offering optimal performance-to-size ratio and outperforming many closed-source competitors.

WizardLM-2 8x22B is Microsoft's latest Mixture-of-Experts model, using 8 expert networks of 22B parameters each. This 176B total parameter architecture delivers competitive performance in knowledge-intensive tasks.

Llama-3-8B-Instruct is Meta's compact dialogue model with 8B parameters, trained on 15T tokens. This instruction-tuned decoder demonstrates competitive performance against larger closed-source models in human evaluations.

Qwen-2.5-72B-Instruct leads the Qwen2.5 series at 72B parameters. This instruction-tuned transformer excels in language understanding, coding, mathematics, reasoning, and multilingual tasks.

A high-performing open embedding model with a large token context window

Ask for feedback on Redux code based the Redux style guide

Java23 API