Project Mirage.
A math tutor built in Venutian. It starts with Srinivasa Ramanujan 500 (SR500), a mathematics language model with no attention layers, for anyone preparing for a competitive exam.
0.44 percent off the 500M target.
Replaced end to end by damped complex resonators.
Token one and token 100,000 cost the same.
Paired with an exact computation core.
What it is
A math tutor that runs on a laptop.
SR500 works a problem through step by step, the way a patient tutor would, instead of producing a bare final answer. It has no attention layers and is built in Venutian, small enough to run on an ordinary laptop.
$ A tank is filled by pipe A in 6 hours and emptied by pipe B in 10 hours. Both are opened together. How long until the tank is full?
Work with rates, not times.
Pipe A fills 1/6 of the tank each hour.
Pipe B empties 1/10 of the tank each hour.
Opened together, the net rate is 1/6 minus 1/10.
Common denominator 30: that is 5/30 minus 3/30, which is 2/30, or 1/15 of the tank each hour.
Filling at 1/15 of the tank per hour, the full tank takes 15 hours.
Check: in 15 hours, pipe A alone would add 15/6 = 2.5 tanks; pipe B alone would remove 15/10 = 1.5 tanks. 2.5 minus 1.5 is 1 tank. Correct.
Answer: 15 hours.An example of SR500's tutoring style. The same patient, step by step voice carries to any problem it is asked.
What replaces attention
Damped complex resonators. In place of attention, every layer carries a bank of damped complex resonators, each with its own decay rate and its own ringing frequency. Every token strikes the forks; the forks ring on and slowly fade.
Linear in length. What a token sees of everything before it is the state of the forks when it arrives. Attention compares every token against every other token, quadratic work that gets heavier the longer the problem runs.
The resonator recurrence is linear and associative, so it solves as a parallel scan in log depth passes instead of one pass per token.
Built in Venutian. SR500 is trained and orchestrated in Venutian, on H100s rented through Modal.com.
- Tokens
- 8,192
- SR500 vs. attention
- 75.4x
- SR500 decode state
- 282 KB, fixed
Measured results
Speed and accuracy, with sources.
Speed
Attention's comparisons multiply with the square of the length. The parallel scan's work only grows in log depth, so the advantage keeps widening the longer the problem runs.
Measured on CPU, forward only: resonator scan vs. attention, by context length.
Show as table
| Context length | Measured |
|---|---|
| 512 tokens | 5.6x faster than attention |
| 2,048 tokens | 28.2x faster than attention |
| 4,096 tokens | 45.2x faster than attention |
| 8,192 tokens | 75.4x faster than attention |
GSM8K
Paired with its exact computation core as a verifier, SR500 scores 92.66 percent on GSM8K, run against small specialist models and against frontier labs' own flagships.
| Model | GSM8K | Notes |
|---|---|---|
| SR500 (Venutian, 502M) | 92.66% | With its exact computation core as verifier |
| Small specialist models | ||
| Phi-3-mini (3.8B) | 86.9% | Microsoft's own reported figure, 8-shot chain of thought |
| Llama 3 8B Instruct | 79.6% | Meta's own reported figure, 8-shot chain of thought |
| Frontier labs' flagships | ||
| GPT-4 (original) | 92.0% | OpenAI's technical report, 5-shot chain of thought |
| Claude 3 Opus | 95.0% | Anthropic's model card, 0-shot chain of thought |
| Gemini Ultra | 94.4% | Google's technical report, 11-shot |
How to read these figures
Every competitor figure comes from that lab's own technical report or model card, under its own evaluation conditions, shot count and prompting included. This is a comparison of published numbers, not one controlled study.
SR500's score is its result with its exact computation core doing the arithmetic, a real part of how it runs. SR500 is a 502 million parameter model small enough to run on a laptop, and on this benchmark it is competitive with models many times its size.
Exact arithmetic
The exact computation core.
SR500 pairs its resonator reasoning with an exact computation core that carries out the arithmetic itself and verifies every result. The model never has to hold a long division in its head.
Reasoning in the resonator, computation in the core: that pairing is what the 92.66 percent GSM8K with verifier figure reflects.
Right to education
Access to coaching is uneven.
Every year, millions of students sit for competitive exams hoping to change their own lives. For a great many of them, the deciding factor is whether their family can afford coaching.
SR500 exists to narrow that gap. It puts a patient, capable tutor within reach of anyone with a computer.
Once installed, SR500 runs locally. It keeps working through a power cut, a network outage or a village with no reliable broadband, with no data cost between a student and the next explanation.
It is there at two in the morning before an exam, when the institute is closed and there is no one else to ask.
Built for
- Students preparing for engineering and medical entrance exams, JEE and NEET among them, who cannot afford a year of coaching.
- Anyone studying for a competitive exam with a serious mathematics or quantitative section, from board exams to postgraduate and civil-service entrance tests.
- Students in towns and villages with unreliable or expensive internet, since SR500 is built to run locally and needs no connection.
- Self-learners who need a patient, step by step explanation at whatever hour they are stuck, including outside coaching institute hours.
- Teachers and schools who want a supplementary tool for students who cannot access private tuition.
Project Mirage
Get notified.
SR500 ships with Venutian 1.0 and the marketplace.