Overview & Capability
High-throughput Mixture-of-Experts (MoE) foundation model activating 37B parameters per token with Multi-head Latent Attention (MLA).
Key Highlights
- 60 tokens/second inference throughput on modern clusters
- 128k context window with perfect retrieval accuracy
- Leading open-weight benchmark on MMLU and coding datasets
Robbie's Independent Take
Verified: 2026-10-11
Ideal for latency-critical agent loops where fast token generation directly reduces user wait times. Benchmark output quality on your specific task before swapping out larger frontier reasoning models.
Alternatives & Replaces
- Llama 3.3 70B
- GPT-4o