Celeris-1
#100 in Frontier LLMsCeleris Labs · v1 · 3× · last seen Jul 28, 2026
21
Momentum
Celeris-1 is the first language model from AI lab Celeris Labs in San Francisco, unveiled on July 22/27, 2026. It is based on a diffusion-based Inference architecture instead of classical autoregressive Token-by-Token generation and produces responses in parallel rather than sequentially. The model is offered exclusively through a proprietary, OpenAI-compatible API (weights not publicly available) and achieves near GPT-5 level performance with response times up to 15 times faster (p50 latency approximately 157-158ms), according to the provider.
Momentum trend
01.05.30.07.
Features
| Key Benchmark (%) | 75.9% on MMLU-Pro (vs. 78% GPT-5-mini, 81% GPT-5) |
| Context Window (Tokens) | 8,192 tokens |
| License | Proprietary, API access only; model weights and full architecture not public |
| Multimodality | No multimodality specified; text-only language model per documentation |
| Platform | OpenAI-compatible cloud API (inference.celeris.ai), streaming support, AWS/GCP VPC deployment for enterprise |
| Price per 1M Tokens | $2 input / $6 output per million tokens, metered separately |
| Release Date | July 22, 2026 (release); announcement July 27, 2026 |