> ## Content Index > Fetch the complete content index at: https://www.implicator.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Uno Researchers Report Up to 3× Faster AI Without Sacrificing Quality - URL: https://www.implicator.ai/uno-lossless-ai-speedup/ - Published: 2026-09-07T15:04:21.000Z - Updated: 2026-09-07T15:04:21.000Z - Description: Uno researchers report up to 3× faster AI generation without changing the base model’s output distribution. Their results distinguish server throughput from single-request speed, while a Google researcher challenges a key comparison. The open code gives outside teams a way to test it. - Author: Marcus Schuler - Tags: AI Research Large language models generated tokens up to 3× faster than their own autoregressive base models in tests reported in a September 3 [Uno preprint](https://arxiv.org/abs/2609.04010?ref=implicator.ai), while preserving those models’ output distributions. Uno adds lightweight diffusion adapters that propose several tokens at once, then lets the original model check them, a design documented in the team’s [released implementation](https://github.com/ifm-ai/uno?ref=implicator.ai). If the results hold, the method could shorten generation for individual requests and increase capacity on servers handling concurrent workloads, although this reporting found no independent verification of its central speedup. Subham Sekhar Sahoo, an Institute of Foundation Models researcher and the paper’s first and corresponding author, and his co-authors built Uno for models that normally produce one token after another. Their tests cover their own model and a version based on open-weight Qwen3-8B. What Changed - Uno researchers report up to 3× faster generation in tested settings while preserving the base language model’s output distribution. - Uno’s paper reports higher server throughput than DiffusionGemma, but slower generation for a single request. - Google researcher Brendan O’Donoghue challenges the competing model’s baseline speed measurement in the paper. - The reported central Uno speedup has not been independently verified in this reporting. Code and checkpoints are available for outside testing. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). ## One model, two pathways A standard autoregressive model predicts the next token from the tokens that came before it. That sequence can leave a graphics processor waiting on repeated transfers of model weights and cached context, especially when one request is being served. For diffusion-adapter training, Uno keeps those autoregressive weights frozen and attaches diffusion adapters. Using the same underlying model, the adapters learn to turn a noisy block into several candidate tokens in parallel before the original autoregressive pathway checks them. The base pathway then applies a rejection correction to those candidate tokens, accepts the longest valid prefix and samples a replacement token when a candidate fails its check. “Lossless” has a narrow mathematical meaning here. The correction preserves the probability distribution of the base model. It does not promise identical text on every run, fewer hallucinations or greater intelligence. The adapters are trained to accelerate an existing model, not improve what that model knows. FREE WEEKDAY MORNING BRIEFING Follow the tests behind faster AI. The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern. Email address Send me tomorrow’s briefing Check your inbox for the confirmation link. About five minutes. No hype. No spam. Unlike conventional speculative decoding, Uno does not require a separately trained draft model. Its proposal and verification pathways also share one architecture and one key-value cache, reducing added memory in the paper’s comparisons. ## What the tests measured The researchers derived the central speed figure from controlled throughput tests on one Nvidia H200 using Nano-vLLM. Each run used 1,024 random input tokens and a fixed effective output of 8,192 tokens. The researchers first measured accepted tokens per forward pass across benchmarks, then imposed those averages on the timing runs. Benchmark scores were measured separately from the clocked server runs and do not supply independent timing evidence. That procedure holds sequence length constant, but it is not direct end-to-end timing of coding agents, tool calls or other production tasks. Released code and checkpoints permit outside testing; they do not by themselves establish production readiness. The authors advertise speedups of up to 3× at the best settings and up to 2× at the largest supported batch. Detailed Uno-Qwen results were about 2.5× at batch one and 1.6× at its largest batch. The team’s own trained model measured about 2.2× at batch one and 1.5× at batch 64. ## A split throughput result In the paper’s main comparison table, Uno reached 5,255 tokens a second in aggregate throughput, compared with 1,136 for [DiffusionGemma](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/?ref=implicator.ai). Per-request results reversed the order: Uno reached 405 tokens a second, while DiffusionGemma reached 836. Know someone who'd find this useful? [✉️ Email it to a friend in one click](mailto:?subject=A%20newsletter%20I%20think%20you%27d%20like&body=This%20is%20one%20of%20maybe%20three%20newsletters%20I%20actually%20read.%20The%20rest%20just%20pile%20up%2C%20unread%2C%20judging%20me.%0A%0AAnd%20yes%2C%20this%20email%20mostly%20wrote%20itself%2C%20which%20is%20a%20little%20on%20the%20nose%20for%20an%20AI%20newsletter.%20Doesn%27t%20make%20it%20wrong.%20implicator.ai%20is%20good.%0A%0ASubscribe%20free%3A%20https%3A%2F%2Fwww.implicator.ai%2Fsubscribe%2F%3Futm%5Fsource%3Dnewsletter%26utm%5Fmedium%3Dforward%26utm%5Fcampaign%3Demail%5Fforward), or they can [subscribe free here](https://www.implicator.ai/subscribe/?utm%5Fsource=newsletter&utm%5Fmedium=forward&utm%5Fcampaign=forward%5Fto%5Fcolleague). The split matters because server capacity and an individual user’s wait are different measurements. Uno’s linear sampler is tuned for aggregate throughput at high concurrency, while its tree sampler uses spare compute to improve single-request speed. Quality comparisons also need a boundary. Uno’s cross-model results put it ahead on many agentic, coding and long-context tests, but not every listed benchmark. On AA-Omniscience, Uno scored 14.3 against DiffusionGemma’s 17.7\. The lossless guarantee applies against Uno’s own autoregressive base, not across different models trained on different data. ## The disputed baseline Brendan O’Donoghue, a Google Research Scientist and author of the DiffusionGemma announcement, challenged its baseline in a [September 6 post](https://x.com/bodonoghue85/status/2096531010766692507?ref=implicator.ai). His H200 test used vLLM, PG19 data, a 1K-input and 8K-output shape, and a fixed 17.56 tokens per forward pass. O’Donoghue measured 2,700 tokens a second at BF16, closest to the paper’s 1,136 figure in precision, and 3,000 at FP8\. As the author of the competing model’s announcement, his test is not an independent replication of Uno or evidence against the lossless method. Different serving engines and input data also limit comparison. O’Donoghue said further vLLM tuning could lift throughput: “The real number is 3k tok/s or higher.” Frequently Asked Questions What does Uno add to a language model? Uno adds lightweight diffusion adapters that propose several tokens at once. The original model checks the candidates, preserving its output distribution without requiring a separate draft model. What does lossless mean here? It means preserving the base model’s probability distribution. It does not guarantee identical text on every run, fewer hallucinations, or greater intelligence. Is Uno faster than DiffusionGemma for a single request? No. The paper’s up-to-3× result applies to tested settings against its own base model. Its main comparison table gives Uno 405 tokens per second for a single request and DiffusionGemma 836, despite Uno’s higher aggregate server throughput. What is disputed about the benchmark? Brendan O’Donoghue reports 2,700 tokens per second for DiffusionGemma at BF16, compared with 1,136 in Uno’s paper. His test used a different serving engine and input data. It challenges the DiffusionGemma baseline rather than replicating Uno. Has Uno’s speedup been independently verified? The central speedup has not been independently verified in this reporting. Released code and checkpoints permit outside testing. The paper’s timing tests used controlled input and output lengths rather than end-to-end production tasks. AI-generated summary, reviewed by an editor. [More on our AI guidelines](https://www.implicator.ai/about/). [Apple's New Mac Studio Targets the Wait Before the First TokenApple introduced a new Mac Studio with M5 Max and M5 Ultra on August 25, 2026, saying the M5 Ultra processed prompts up to four times faster than the M3 Ultra in its July 2026 tests. Neural AcceleratoThe Implicator![](https://www.implicator.ai/content/images/2026/08/20260825-224728-waiting_first_token_v2.webp)](https://www.implicator.ai/apple-mac-studio-m5-ultra-prompt-processing/) [Inception Ships Mercury 2, a Diffusion LLM That Hits 1,009 Tokens Per SecondInception on Tuesday released Mercury 2, a language model built on diffusion architecture that the startup says generates text more than five times faster than comparable systems, Bloomberg reported. The Implicator![](https://www.implicator.ai/content/images/2026/02/2026-02-24-06.42.01-typewriter_burst@2x.webp)](https://www.implicator.ai/inception-ships-mercury-2-a-diffusion-llm-that-hits-1-009-tokens-per-second/) [Inception banks $50M to replace how AI writes codeThree professors bet diffusion beats autoregression for speed. Microsoft, Nvidia back parallel token generation Inception just raised $50 million to prove language models have been generating text wrThe Implicator![](https://www.implicator.ai/content/images/2025/11/2025-11-06-06.28.33-inception_ai_lab_1@2x.webp)](https://www.implicator.ai/inception-banks-50m-to-replace-how-ai-writes-code/)