Proposal efficiency
Investigating how to reduce the overhead of getting useful candidate tokens in front of a target model.
Our research / Your exclusive model
Faster is the goal.
Lossless is the rule.
fix8 stands for Frontier Intelligence eXploration. Our research powers the models we optimize and host. Start with your model, one of ours, or a new model for your objective. We achieve the agreed result, then host that optimized model as a black box for your exclusive use. Every optimization is lossless.
Our working principle
Keep the intelligence.
Change the work required to reach it.
Lossless is a requirement, not an optional mode. Every optimization we develop must preserve the target model’s output fidelity. We change the work required to get the result, without sacrificing the result.
01 / Speculative decoding
A different way
to make progress.
Traditional autoregressive decoding advances token by token. Each step carries a cost, even when the next part of a response is relatively predictable.
Speculative decoding proposes a sequence using a less expensive path, then asks the target model to verify it. Accepted tokens let generation advance further in one verification pass.
The practical challenge is the balance: proposals must be useful enough, and cheap enough, to outweigh the extra work they introduce. That balance changes with the model and workload.
Propose
Generate candidate tokens ahead of the target model.
Verify
Use the target model to determine which candidates can be accepted.
Advance
Move forward on verified work, then repeat.
02 / Current observations
21–53% faster.
A starting point.
Our internal testing has achieved lossless inference speed improvements of 21–53% across Qwen 3, Gemma 4, GPT-OSS, and other models.
Latest internal observation: Gemma 4 has achieved a 50.73% lossless inference speed increase.
That range describes observed improvements across tested configurations. It does not assign a particular percentage to each model or imply that every workload sees the same result.
Performance is contextual. Hardware, model size, execution settings, prompt shape, generation length, and concurrency can all affect the outcome.
We have not published a full reproducible benchmark package here. If you want to discuss the test conditions or explore a comparison on your workload, get in touch .
03 / Inside the advantage
Our methods stay inside.
The performance is yours.
Workload awareness
Exploring when an optimization should adapt, scale back, or make a different execution choice.
Systems and algorithms
Researching new interactions between decoding strategies and the systems that execute them. These methods underpin exclusive hosted model services. You use your optimized model; the algorithms and implementation behind its optimization remain with fix8.
Your model.
Your exclusive service.
Tell us what you want to achieve, whether you already have a model, and how you need to use it. We’ll discuss optimization followed by exclusive hosting, with our methods kept behind the service.
Discuss your model