InternLM

InternLM2.5 and InternLM2 chat models support conversational interactions across multiple languages with a 200K token context window, enabling long-form document analysis and multi-turn dialogue. The models are fine-tuned via supervised fine-tuning (SFT) on instruction-following datasets, allowing them to follow complex user directives while maintaining coherence across extended conversations. This is implemented through standard transformer decoder architecture with rotary position embeddings (RoPE) scaled for long-context handling.

InternLM3 introduces a specialized 'deep thinking mode' that enables the model to perform extended chain-of-thought reasoning for complex mathematical problems, logic puzzles, and multi-step reasoning tasks. This mode works by allowing the model to generate internal reasoning traces before producing final answers, implemented through a two-stage generation process: first generating hidden reasoning tokens (not shown to users), then producing the final response. The architecture uses a modified attention mechanism that allows the model to 'think' without token budget constraints on visible output.

InternLM is expanding into multi-modal capabilities through integration with vision encoders, enabling models to process images alongside text. This is implemented by combining a vision encoder (e.g., CLIP-based) with the language model backbone, where images are encoded to visual tokens and concatenated with text tokens in the input sequence. The model learns to reason about both visual and textual information through instruction-tuning on image-text datasets. This enables applications like image captioning, visual question answering, and document understanding from scanned PDFs.

InternLM provides support for deployment on NPUs (Neural Processing Units) such as Huawei Ascend, enabling efficient inference on edge devices and specialized hardware. This is implemented through model quantization (int8, int4) and NPU-specific optimization passes that convert standard transformer operations to NPU-native operations. The framework handles model compilation, memory management, and operator fusion for NPU targets. This enables deployment of InternLM models on edge devices with significantly reduced latency and power consumption compared to GPU inference.

InternLM provides tools for converting models between different formats and frameworks, including conversion to ONNX, TensorRT, and other inference-optimized formats. The conversion pipeline handles weight transformation, operator mapping, and format-specific optimizations. This enables deployment of InternLM models in diverse inference environments (ONNX Runtime, TensorRT, TVM, etc.) without retraining. The tools also support quantization during conversion, enabling efficient deployment on resource-constrained devices.

InternLM2.5 and InternLM2 models support structured function calling through a schema-based approach where tools are defined as JSON schemas and the model learns to emit properly formatted tool calls within its generation. The implementation uses a special token vocabulary for tool invocation and integrates with frameworks like LMDeploy and SGLang that parse model outputs and route calls to registered functions. This enables agentic workflows where the model can autonomously decide when and how to use external tools (APIs, calculators, databases) based on user intent.

InternLM models are trained on large code corpora and support code generation, completion, and understanding tasks across 40+ programming languages. The models learn to generate syntactically correct code through exposure to high-quality open-source repositories during pretraining. Code understanding is enhanced through instruction-tuning on code-related tasks (debugging, explanation, optimization). The architecture uses standard transformer attention but benefits from code-specific tokenization that preserves syntax structure, enabling better handling of indentation and bracket matching.

InternLM2.5 extends context handling to 1 million tokens through continued pretraining with specialized position interpolation techniques and efficient attention mechanisms. The implementation uses a combination of RoPE scaling, grouped-query attention (GQA) for memory efficiency, and training on synthetic long-context data to enable processing of entire books, codebases, or document collections in a single context window. This is achieved without catastrophic forgetting of the base 200K capability through careful curriculum learning during continued pretraining.

InternLM provides XTuner, a specialized fine-tuning framework that enables efficient adaptation of base models to specific domains or tasks through low-rank adaptation (LoRA), quantized LoRA (QLoRA), and full fine-tuning. XTuner handles the complete fine-tuning pipeline including data loading, training loop management, gradient accumulation, and model checkpointing. The framework integrates with InternLM's training infrastructure and supports both single-GPU and distributed training, making it accessible for teams without massive compute budgets.

InternLM integrates with LMDeploy, a specialized inference toolkit that optimizes model serving through techniques including key-value (KV) cache quantization, continuous batching, and dynamic shape inference. LMDeploy compiles models to an optimized intermediate representation (IR) and uses a custom CUDA kernel library for efficient attention computation. The toolkit supports multiple deployment backends (local, Docker, Kubernetes) and provides REST/gRPC APIs for model serving, enabling production-grade inference with 2-3x throughput improvement over naive implementations.

InternLM provides an agent framework that enables models to autonomously plan and execute multi-step workflows using multiple tools. The agent system uses a planning-execution loop where the model generates action plans (decomposing user intent into tool calls), executes tools, observes results, and refines plans based on feedback. This is implemented through a specialized prompt template that guides the model through reasoning, tool selection, and result interpretation. The framework supports both sequential and parallel tool execution, with built-in error handling and retry logic.

InternLM provides reward models (separate from base models) trained to score model outputs on quality dimensions (helpfulness, harmlessness, honesty). These reward models are used in RLHF pipelines to fine-tune base models based on human preference data. The reward models are trained via supervised learning on preference pairs (chosen vs rejected responses) and learn to assign scalar scores that correlate with human judgments. This enables iterative improvement of chat models through preference optimization algorithms like DPO (Direct Preference Optimization) or PPO (Proximal Policy Optimization).

InternLM provides a web-based demo interface built with Gradio or Streamlit that enables interactive exploration of model capabilities without coding. The interface supports real-time chat, parameter adjustment (temperature, top-p, max tokens), and visualization of model behavior. The demo can be deployed locally or on cloud platforms, making it accessible for non-technical users to test model outputs. The interface integrates with LMDeploy for efficient inference, enabling responsive interactions even with large models.