Most speed improvements in large language models come after the fact. A lab trains a model, then an inference provider figures out how to make it run faster. Celeris is taking the opposite approach: building models where throughput is a design constraint from the start, not an optimization layered on top.

The company's thesis is straightforward. If you want intelligence delivered per unit of time rather than per parameter count, you need to rethink both how models learn and how they generate. Celeris is beginning with diffusion, the same class of models that powers image generation, and adapting it for language in ways that go beyond bolting parallel decoding onto an autoregressive backbone.

Why Autoregressive Decoding Is the Bottleneck

Every mainstream language model generates text one token at a time. This sequential process means that a 500-token response requires 500 serial steps, each waiting on the previous one. The approach produces coherent language, but it caps throughput at whatever the slowest serial step allows.

Diffusion models in vision already demonstrated that parallel generation works. Instead of producing pixels sequentially, diffusion generates entire images in parallel through iterative refinement. Celeris sees no reason language cannot follow the same logic, and they are not alone in that observation.

What the Research Shows

NVIDIA's Fast-dLLM study demonstrated that decoding many tokens per step on existing diffusion models yields up to a 27.6x throughput improvement with minimal accuracy loss. The gains came from training-free acceleration alone, meaning the underlying models were not retrained for the experiment. That result establishes the ceiling for what diffusion parallelism can deliver without architectural changes.

Parallel decoding also solves a problem autoregressive models cannot. The "reversal curse" shows that models trained on "A is B" fail to infer "B is A." GPT-4 answers 79% of forward questions about celebrity relationships correctly but only 33% of reversed ones. LLaDA, an 8B-parameter diffusion model that attends to the full sequence at every step, breaks this pattern and surpasses GPT-4o on reversal reasoning tasks. Bidirectional context means the model sees the entire prompt simultaneously rather than processing it in fixed left-to-right order.

Diffusion-of-Thought, presented at NeurIPS 2024, demonstrated a third advantage: refinement. An autoregressive model commits to every token permanently once generated. A diffusion model can revise earlier predictions as the rest of the response takes shape. In the study, a small diffusion model self-corrected its reasoning mid-generation and outperformed a much larger autoregressive model on math tasks, both in accuracy and efficiency. The number of refinement steps acts as a natural dial between latency and reasoning quality, something autoregressive architectures lack entirely.

What Celeris Is Building Differently

Celeris is not simply applying diffusion to text generation. The company is developing hybrid approaches where sequential and parallel generation coexist inside a single model. The model spends sequential compute only where language demands it, like maintaining grammatical coherence across long dependencies, and decodes everything else in parallel.

The critical difference from prior hybrid methods is that this design is baked into both the architecture and the training objective. Bolting parallel decoding onto an autoregressive model at inference time produces artifacts and coherence problems. Training a model from scratch with parallelism as a first-class citizen avoids those issues, but it requires rethinking how language models are trained in the first place.

Celeris is also investing in neural architecture search and automated research loops to explore the space of architectures, training objectives, and inference strategies optimized for latency and intelligence density. The goal is not a single fast model but a research engine that continuously discovers better ways to deliver frontier-level capability at microsecond-to-millisecond timescales.

What Fast Language Models Unlock

The applications Celeris envisions are constrained by current latency, not by model capability. Voice interfaces that feel instantaneous require language generation fast enough to match human conversation speed without awkward pauses. Agents operating in tight feedback loops need to reason and act in real time. Scientific and engineering tools that keep pace with human thought demand sub-second response times on complex queries.

Today, these use cases require either small, fast models with limited reasoning ability, or large, slow models that introduce unacceptable delay. Celeris wants to collapse that tradeoff entirely.

The Harder Path

The company acknowledges they are choosing a harder path than incremental optimization. Rethinking training and architecture carries more risk than making existing approaches faster. But Celeris is betting that the payoff in capability per second will be decisive as language models become embedded in real-time applications across every industry.

The company has not announced specific model releases or benchmarks yet, but says first models are coming soon. For teams building applications where latency directly impacts user experience or agent performance, the approach Celeris is pursuing could reshape what is possible with language models in production.