Frontier.fast is a new open-source competition platform that invites anyone to submit code improvements to make AI model inference faster.
Each submission is built from source, tested on dedicated hardware against a live leaderboard, and only accepted if it is genuinely faster while maintaining identical model behavior.
The platform combines local iteration with ranked verification to separate real speed gains from unverified claims, addressing a key bottleneck in AI deployment.
What happened
A new platform called frontier.fast has launched an open competition where anyone can submit code patches to make large language model inference faster. Submissions are tested on dedicated hardware against the current record, and only accepted if they are genuinely faster and the model still produces identical outputs.
Why it matters
Inference speed—how quickly an AI model produces answers—is a critical bottleneck for real-world AI applications. By creating a transparent, reproducible leaderboard with standardized benchmarks across different GPUs and model families, the platform makes performance gains measurable and verifiable, rather than claimed in marketing materials.
What to watch
The scoring formula weights decode speed (65%), prefill (20%), and time-to-first-token latency (15%), with correctness verified by perplexity equivalence within 0.5%. Submissions are free; the platform is currently seeking hardware and funding to expand from its initial dedicated GPU boxes (a DGX Spark GB10, a Radeon AI PRO R9700, and more).
Frontier.fast is a leaderboard-driven competition designed to measurably accelerate AI model inference across different hardware and runtime environments. The platform operates on a simple principle: anyone can submit a code patch (a commit against a pinned engine source) that claims to make a model run faster, and the patch is automatically built from source, run on dedicated hardware in a controlled session against the current record, and either accepted or rejected based on whether it is genuinely faster and produces identical outputs.
The measurement and verification process is the core innovation. Each submission runs alternately in a single session on the same machine as the current record holder, eliminating hardware drift as a confounding factor. Speed is scored using a weighted formula—decode speed (the rate at which the model generates subsequent tokens) accounts for 65% of the score, prefill (the initial computation on the input prompt) accounts for 20%, and time-to-first-token (latency until the model produces its first output) accounts for 15%. Correctness is validated by checking that the submission's perplexity equivalence is within 0.5% of the baseline model. All results—both successful improvements and failures—are published with their supporting evidence and the commit that produced them.
The submission workflow separates fast local iteration from trusted ranked verification. Contributors first sign in with GitHub and choose a frozen track (which pins a specific model, GPU, quantization scheme, runtime, and scoring rules). They then clone the engine source and iterate locally, using quick estimates to guide their optimization work. Only when they have a coherent change do they submit it for official verification. The platform currently runs on dedicated GPU hardware including a DGX Spark GB10 and a Radeon AI PRO R9700, with more hardware planned as the competition grows. All submissions are free and will remain free; the platform is seeking hardware and funding donations to scale the infrastructure.
Inference optimization has emerged as a critical frontier in AI deployment. While model pretraining and fine-tuning capture most research attention, the speed at which a deployed model produces outputs directly affects cost-per-inference, user experience, and real-world viability. Frontier.fast addresses a measurement credibility problem: claims of speedup are often opaque, hardware-specific, or difficult to reproduce independently. By pinning the model, GPU, engine, and benchmark window while leaving optimization approach free, the platform creates a controlled environment where improvements can be verified. The scoring formula (65% decode, 20% prefill, 15% first-token latency) reflects different usage patterns—decode speed matters most for long-output tasks like code generation, while latency-sensitive applications prioritize first-token time. The requirement that faster submissions maintain "behavioural identity" (perplexity equivalence within 0.5%) guards against optimization tricks that degrade output quality. The platform's reliance on dedicated hardware running around the clock and its openness to any contributor model this as an infrastructure investment rather than a one-time benchmark.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Silicon Motion announced a private placement of US$1 billion in aggregate principal amount of 0% convertible s…
Sundar Pichai, CEO of Alphabet and Google, announced on August 11 that the Gemini app's monthly active users (…
An AI system generated a research draft that strengthened a mathematical bound related to the Riemann hypothes…
Researchers published findings showing AMIE (Video), a Gemini-based AI system, performed at or above the level…
After former lead writer Stella Sacco posted on Bluesky that Saber replaced her with ChatGPT midway through de…
OpenAI's ChatGPT crossed 1 billion monthly users some time ago and hit 1 billion weekly users in July, accordi…
The AI news that matters, in one minute each morning.
Sign up free