A startup called Celeris has released Celeris-1 Magnus, a hybrid diffusion language model built for agentic work that it says runs faster than leading autoregressive models on tool-heavy tasks. The model is derived from Qwen3.8-27B, exposes an OpenAI-compatible API, and is available to call today with no waitlist.
Try It: Point Your Agent at a Faster Endpoint
Celeris ships an OpenAI-compatible endpoint at inference.celeris.ai, so you can keep the agent code you already wrote and swap the base URL and key. Create a key at the console, point an existing tool-calling loop (a coding subagent, a customer-support agent, a retrieval agent) at Magnus, and measure end-to-end task latency against your current model. Because it generates and refines multiple tokens at once rather than strictly left to right, the payoff shows up most on multi-step tool loops, not single short completions. There is an optional enable_thinking flag that trades a few seconds for higher accuracy.
Why It Matters for Creators
Most speed-versus-quality tradeoffs in agent building come down to picking a smaller model and hoping it holds up. Diffusion-style decoding is a different lever: it attacks the sequential-token latency floor directly, which is exactly what slows down agents that make many tool calls in a row. If the approach holds outside vendor benchmarks, builders get a way to make existing agent workflows feel responsive without dropping to a weaker model. It also lands the same week that Inception Labs shipped its own diffusion LLM, Mercury 2.5, signaling that diffusion models are becoming a real category for builders rather than a research curiosity.
Key Details
Architecture: Hybrid diffusion model derived from Qwen3.8-27B, optimized for agentic tool-use workloads.
Context window: 131,072 tokens.
Speed: Celeris claims over 500 tokens per second in many scenarios.
Benchmark (vendor-reported): On the τ³-bench banking suite, Celeris reports 41.2% task completion at a 55-second median, against GPT-5.6-sol at 38.1% and 79 seconds. Magnus is now indexed on Artificial Analysis for independent tracking, and third-party writeups such as explainx's breakdown flag that these numbers are vendor-published and worth reproducing on your own stack.
Access: OpenAI-compatible API, billed per token, keys live immediately per the company's launch announcement.
What to Do Next
Treat the benchmark edge as a hypothesis, not a guarantee. Reproduce the τ³-bench numbers on a workload that looks like yours before wiring Magnus into anything you ship. If the latency win holds on your own tool loops, it is a low-risk swap because the API surface matches what you already use.
