Anthropic released Claude 4 Opus with a 200K token context window and improved agentic reasoning. On the same day, the ZAYA project released an 8-billion-parameter model scoring 78.5 on MMLU — ahead of Llama 3 70B at 75.2. The two releases frame the same market tension: scale at the frontier versus efficiency at the edge.
Context: 200K tokens, up from 100K in Claude 3. Targets long-document and code-base agent workflows without chunking.
Access: Available via Anthropic API with tiered pricing.
Strategic angle: Positions Anthropic for enterprise agents that need to hold entire codebases or contracts in working memory.
Model: 8B parameters, trained on 2T tokens, optimized for 4-bit quantization.
Benchmark: MMLU 78.5, surpassing Llama 3 70B (75.2).
Implication: On-device and edge deployment of high-quality reasoning becomes feasible without cloud dependency.
Model: Phi-3-mini (3.8B parameters) runs locally on Windows 11 devices.
Use cases: Real-time document summarization and code suggestions without cloud data egress.
Signal: Major platform vendors are baking small models into the OS layer, not just the API layer.