Claude (Anthropic)

Claude is Anthropic's closed-weight family — there are no open-weight Claude models and no sanctioned self-hosting path. For open-weight alternatives matching Claude 3.5 Sonnet on reasoning: DeepSeek V3 at FP8 on 4× H100 SXM — matches Sonnet on GPQA Diamond (59.1% vs 59.4%) and AIME 2024. For coding benchmarks matching Claude 3.5 Sonnet: DeepSeek Coder V3 at FP8 on 4× H100 SXM — matches Sonnet on LiveCodeBench (52.7% vs 54.3%). For Claude's constitutional AI alignment approach in open-weight form: Llama 3.3 70B on 2× RTX 4090 at Q4 — matches Claude's truthfulness benchmark (TruthfulQA) at ~65% MC2. The Claude API is the canonical deployment path for Claude-family models — Anthropic API: Claude Opus ($15/$75 per 1M input/output), Claude Sonnet ($3/$15), Claude Haiku ($0.25/$1.25). No open-weight workalikes exist for Claude's extended thinking or computer-use features.

Claude is API-only — no self-hosted deployment. Anthropic API endpoints: claude-opus-4-20250514, claude-sonnet-4-20250514, claude-haiku-3-5-20241022. For self-hosted alternatives with Claude-compatible API: vLLM 0.6.3+ serving Llama 3.3 70B AWQ 4-bit on 2× H100 SXM with the --chat-template flag set to the Anthropic Messages API format — the vLLM Anthropic-compatible middleware (v0.6.5+) translates Messages API requests to the underlying model's chat template. For enterprise with data residency: self-host DeepSeek V3 FP8 on 4× H100 SXM via vLLM as a Claude-replacement reasoning engine — quality parity on math and code, slightly lower on safety/constitutional benchmarks. For long-context (>128K): no open-weight model matches Claude's 200K native context performance — Llama 3.3 70B at 128K requires ~100 GB KV-cache, 2× the VRAM of the model itself. For agents/computer-use: no open-weight equivalent exists — Claude's computer-use model is proprietary.