技术栈

Most capable Claude model — complex reasoning, long-context analysis, agentic tasks.

Best balance of intelligence and speed for production workloads.

Fastest and lowest-cost Claude model for high-volume tasks.

AI-native CLI for agentic software engineering — reads, writes, and runs code autonomously.

Open protocol connecting AI assistants to external tools, data sources, and services.

Build, orchestrate, and deploy multi-agent systems powered by Claude.

Top-tier reasoning model with 128K context — Mistral's flagship for enterprise tasks.

Cost-efficient multimodal model — text and image understanding.

Apache 2.0 multilingual model — EU-sovereign deployments, 128K context.

Code generation specialist — 80+ languages, fill-in-the-middle, 32K context.

Frontier vision-language model — document analysis, chart reading, 128K context.

High-quality text embeddings for RAG and semantic search.

Train and own frontier AI model weights outright — no API rental, full data sovereignty.

Enterprise AI assistant — SSO, audit logs, EU data residency, web search, document upload.

Meta's flagship open-weight model — Apache 2.0, matches GPT-4 on many benchmarks at fraction of cost.

Lightweight Llama models for mobile, edge, and on-device inference.

Vision-language Llama models — image understanding, document analysis.

Google's open-weight family — Apache 2.0, strong reasoning, multilingual, edge-to-server range.

MIT-licensed reasoning specialist — outperforms models 3× larger on math and coding.

Edge-optimised reasoning model — 3.8B parameters, strong instruction following on constrained hardware.

Alibaba's Apache 2.0 multilingual family — exceptional Chinese/English, strong math, full size range.

State-of-the-art open-source code generation — rivals GPT-4o on coding benchmarks.

MIT-licensed reasoning specialist with chain-of-thought — matches o1 on math and science tasks.

671B MoE open-weight general model — top open-source benchmark scores across all categories.

TII's Apache 2.0 family — strong multilingual performance, designed for EU/MENA sovereign deployments.

Ultra-compact models for on-device and browser inference — Apache 2.0, efficiency benchmark.

One-command local model serving — runs Llama, Mistral, Gemma and 100+ models on any hardware.

High-throughput production LLM serving — PagedAttention, continuous batching, OpenAI-compatible.

Hugging Face's production inference server — tensor parallelism, quantization, streaming.

CPU/GPU inference in C++ — GGUF format, runs on Apple Silicon, NVIDIA, AMD, CPU-only.

Desktop GUI for discovering, downloading, and running local LLMs — OpenAI-compatible server.

Run Transformers in the browser and Node.js — ONNX-based, no server required.

Cross-platform optimised inference — CPU, GPU, mobile, browser, WASM support.

Universal LLM API proxy — call 100+ models with OpenAI format, load balancing, fallbacks.

2× faster fine-tuning, 70% less VRAM — LoRA and QLoRA for Llama, Mistral, Qwen, Gemma.

Production fine-tuning framework — YAML config, LoRA/QLoRA/full, multi-GPU, Flash Attention.

Fine-tune 100+ LLMs with a web UI or CLI — SFT, DPO, GRPO, LoRA, QLoRA.

PyTorch-native fine-tuning library — recipe-based, minimal dependencies, full control.

Parameter-Efficient Fine-Tuning — LoRA, QLoRA, IA³, AdaLoRA, Prefix Tuning.

Transformer Reinforcement Learning — SFT, DPO, GRPO, PPO, ORPO for alignment training.

ZeRO optimizer for large model training — 10× throughput, trillion-parameter scale.

One-line multi-GPU and TPU training — no code changes, FSDP and DeepSpeed integration.

NVIDIA's large-scale pre-training framework — tensor/pipeline/sequence parallelism.

900K+ models, 100K+ datasets, and Spaces — the de facto standard for AI artifact sharing.

Core model library — load, run, and fine-tune any model in PyTorch, TensorFlow, or JAX.

100K+ datasets with streaming, arrow-based loading, and one-line preprocessing.

Managed dedicated or serverless model deployment — auto-scaling, private endpoints.

No-code fine-tuning for LLMs and other models — SFT, DPO, classification, NER.

Host Gradio and Streamlit ML demos — free tier available, GPU-enabled options.

Standardised metrics library — BLEU, ROUGE, accuracy, F1, and 100+ custom metrics.

Programmatic Hub access — upload models, create repos, manage tokens, search.

LLM application framework — chains, agents, RAG, tool use, memory.

Data framework for LLM apps — ingestion, indexing, querying over any data source.

Production NLP pipeline framework — RAG, document search, question answering.

Declarative LLM programming — optimise prompts and weights automatically.

Structured output extraction — Pydantic schemas from any LLM, with validation and retries.

Enterprise LLM orchestration for .NET, Python, Java — plugins, planners, memory.

Role-based multi-agent orchestration — agents collaborate with defined roles and goals.

Microsoft's multi-agent conversation framework — async agents, human-in-the-loop.

Minimal agentic framework — code-first agents that write and execute Python, 1000-line core.

Rust-based vector search — on-prem friendly, filterable, sparse+dense hybrid search.

GraphQL API vector database — multi-tenancy, hybrid search, generative search.

Local-first open-source vector database — Python-native, zero infrastructure required.

Distributed vector search for billion-scale data — HNSW, IVF, GPU acceleration.

Vector similarity search extension for PostgreSQL — no separate infrastructure needed.

Managed cloud vector database — serverless tier, namespaces, metadata filtering.

Open-source humanoid robot for research and industry — Apache 2.0, ROS2, Python SDK.

Robot Operating System 2 — real-time communication, sensor fusion, navigation stack.

Open-source robot learning — imitation learning, reinforcement learning, pre-trained policies.

Robot simulation and deployment platform — synthetic data generation, physics simulation.

Computer vision library — 2500+ algorithms, real-time image processing, widely deployed.

Segment Anything Model 2 — real-time video and image segmentation, zero-shot.

Real-time object detection — fastest production-grade detector, ONNX/CoreML export.

Managed foundation model APIs on AWS — Claude, Llama, Mistral, Titan, Stable Diffusion.

Microsoft's enterprise AI platform — model catalog, fine-tuning, responsible AI tools.

GCP's unified AI/ML platform — Gemini, model garden, AutoML, feature store.

Run AI models at the edge globally — Workers AI, 100+ models, serverless inference.

LLM observability and tracing — log runs, compare prompts, regression testing.

ML experiment tracking, visualisation, and hyperparameter sweeps — industry standard.

ML lifecycle management — experiment tracking, model registry, deployment.

Inference metrics collection and dashboards — latency, throughput, error rates.

LLM evaluation and monitoring — hallucination detection, embeddings visualisation, drift.

RAG evaluation framework — faithfulness, answer relevancy, context precision metrics.