Post-training platform — in alpha testing
Frontier capability. A fraction of the footprint.
Nexus turns a task description and a handful of your examples into a small specialist language model — a core — built to match far larger models on the one job you need, and to run wherever your data lives.
Apex — frontier-scale teacher, hundreds of billions of parameters
Core — your specialist, 1–8 billion
00 The name
-
I
Apex
The teacher.
Frontier-scale open-weight models at the peak of general capability, and of cost. Hundreds of billions of parameters, most of them irrelevant to your task.
-
II
Nexus
The platform.
Where your task, your examples, the teachers and a frozen test set meet. Synthesis, distillation, reinforcement and evaluation, automated end to end.
-
III
Core
The result.
A specialist of one to eight billion parameters, cast for a single job. Weights you own, in open formats, on hardware you choose.
01 The problem
Most AI work is narrow. Most AI spend isn't.
Extract these fields. Route this ticket. Check this clause against that policy. The language tasks that run millions of times a day are narrow and well defined — yet they are routinely sent to the largest general-purpose models available, which bill every token for a knowledge of poetry, chemistry and forty programming languages the task will never touch.
- Cost
- Spend grows linearly with usage, forever. There is no volume at which a rented generalist becomes cheap.
- Latency
- A round trip to a frontier model is measured in seconds. Many products need milliseconds.
- Exposure
- Every request ships your customers' data to someone else's servers.
- Drift
- The model behind the API changes or retires on someone else's schedule, and your product's behavior changes with it.
- Energy
- Running hundreds of billions of parameters to fill in a form is a poor use of a grid.
Research has shown again and again that a small model trained for one task can match a far larger generalist on that task. What has been missing is a way to get one without an ML research team.
02 The platform
Describe the task. Nexus casts the core.
Specify
Describe the job in plain language and add a few dozen examples. Nexus compiles a task specification — schema, rubric, edge cases — and freezes a held-out test set before any training begins.
Synthesize
Large open-weight teacher models generate, critique and filter hundreds of millions of tokens of training data covering your task space, including the rare and adversarial cases your examples missed.
Distill
Candidate base models from one to eight billion parameters are fine-tuned on the synthesized corpus. Nexus keeps the smallest one that clears your bar.
Reinforce
Reinforcement learning against verifiable rewards — schema validity, programmatic checks, rubric-based judges — hardens the core on the long tail where fine-tuning alone falls short.
Prove
Every candidate is evaluated head to head against its teacher on the frozen test set, a regression suite and adversarial probes. A core ships with its evidence, or it doesn't ship.
03 What you own
Weights, not an API key.
- Parameters
- 1–8 billion
- Formats
- safetensors · GGUF · ONNX, with 4- and 8-bit quantized builds
- Runs on
- a single data-center GPU, a CPU server, or a laptop-class device, depending on size
- Serving
- OpenAI-compatible endpoints through vLLM or llama.cpp, or embedded in your application
- Ships with
- evaluation report · model card · reproducible training manifest with full data lineage
- Ownership
- the weights are yours; base-model licenses are open and disclosed up front
- Your data
- used for your core only, never to train anyone else's
04 Where the compute goes
We spend compute once, so you can stop spending it forever.
A core is cheap to run because it was expensive to make. Casting one is a burst of heavy accelerator work: serving teacher models with hundreds of billions of parameters to write the training data, sweeping candidate students, running reinforcement-learning rollouts by the million, and scoring every checkpoint against its teacher.
| Stage | GPU-hours | Share |
|---|---|---|
| Synthesize teacher inference | 400–1,500 | 30% |
| Distill fine-tuning sweeps | 150–500 | 10% |
| Reinforce RL rollouts and updates | 800–2,500 | 53% |
| Prove evaluation against the teacher | 100–300 | 7% |
| Total GPU-hours per core | ≈ 1,500–4,800 | 100% |
Planning estimates from our own modeling for a 4–8B-parameter specialist — not measurements. Actuals vary with task, corpus size and base model. Platform research — ablations on synthesis strategy, reward design and base-model selection — multiplies this several-fold before the first production core ships.
Accelerator capacity is the single largest input to this company — and every hour of it is pointed at making AI cheaper to run.
05 Why it matters
Owning a model shouldn't require a research lab.
Private by construction
A core runs where your data lives: your cloud, your server room, your device. Nothing leaves at inference time, because there is nowhere for it to go.
Lighter on the grid
A four-billion-parameter specialist does a small fraction of the arithmetic per token of a frontier-scale generalist. Multiply that by every routine request an organization makes.
Yours to keep
No deprecation notices, no silent behavior changes, no per-token meter. The model changes when you change it.
Within reach
Hospitals, public agencies and small companies have narrow, high-value language tasks and no ML team. Nexus is built for them.
06 Target tasks
Built for the jobs that run a million times.
Illustrative target tasks
Extract
Structured extraction
Invoices, claims, lab reports and contracts into schema-valid JSON.
Route
Triage and routing
Tickets, emails and alerts into category, priority and next action.
Check
Policy and compliance checks
Does this document meet that policy, with the clause cited.
Answer
Bounded-domain Q&A
Product manuals and internal procedures, answered from your own corpus.
Query
Text to query
Natural language into SQL, log queries or your internal DSL, for your schema.
Edge
On-device assistants
Offline, private, low-latency help in the field, the clinic or the kiosk.
07 Status
We're early. That's the opportunity.
Nexus is in alpha testing. We are looking for a small number of design partners with one narrow, high-volume language task and a clear bar for success.
Tell us three things: the task, the volume, and what "good enough" means. We'll tell you honestly whether a core is the right tool.