

celeris-1
#199 v Frontier jazykové modelyUnknown · v1 · od 22. Juli 2026 (offiziell angekündigt 27. Juli 2026) · 4× · naposledy 17. 8. 2026
Celeris-1 is the debut model of Celeris Labs (founded by Tom Hamer and Jesse Clark, previously Marqo), announced on July 22/27, 2026. Unlike conventional autoregressive LLMs, it uses a diffusion architecture that generates multiple tokens in parallel, delivering near-GPT-5-level intelligence with up to 15x faster response times according to the vendor. It is accessible via an OpenAI-compatible API, is proprietary (weights not open), has a combined prompt+completion context window of 131,072 tokens, and scores 75.9% on MMLU-Pro at a p50 latency of 158ms and roughly 1,600-2,000 output tokens per second.
Vlastnosti
| Key Benchmark (%) | 75.9% on MMLU-Pro (5-shot, chain-of-thought, no reasoning budget) |
| Context Window (Tokens) | 131,072 tokens (combined prompt + completion) |
| License | Proprietary, model weights not publicly available |
| Multimodality | Text and image input, text-only output |
| Platform | OpenAI-compatible API (inference.celeris.ai) |
| Price per 1M Tokens | $2 per 1M input tokens / $6 per 1M output tokens |
| Release Date | July 22, 2026 |