Models
Specialized, not general-purpose.
REMIT does not run a chatbot. Each workflow gets a small model trained on that workflow and measured on your own past cases, self-hosted on open weights inside your perimeter.
Trained on your workflow.
A model that knows one job well beats a general one on that job. Each solution ships with its own.
Open weights, no lock-in.
Self-hosted on your hardware. Nothing depends on a vendor API, and you own the right to run the weights.
The guarantee is not the model.
The accuracy standard is measured on your cases, so the model can change and the proof survives.
The model base
Sized to the job
Small where the task is choosing from a fixed list. Larger only for the minority of hand-offs, where reading evidence carefully is worth seconds per case.
| Product | Task | Model | Why this size |
|---|---|---|---|
| FOLIO | Find the relevant policy passages | Small retrieval and re-ranking models | Runs on a branch-class machine, per query. |
| FOLIO | Read documents into a case record | Small model, fixed output format | Narrow task. Facts come from the documents, not from memory. |
| SIFT | Choose from your list of outcomes | Small model, fine-tuned per workflow | Picking from a short list needs no world knowledge. Smallest size that meets your standard. |
| BENCH | Deep search over the full corpus | Larger retrieval models | Higher recall for the hard cases. Runs on the core server. |
| BENCH | Suggest, with citations, on hand-offs | Core model, sized to the workflow | Hand-offs are a minority, so seconds per case is acceptable. |
BOUNDS has no model of its own. It is the same measurement procedure for every workflow.
Core tier
A choice, not a lock-in
Selected per estate. Solutions move between tiers and are re-measured.
Lean
Single workflow
Lowest hardware, fastest to deploy.
Standard
Most regulated decision work
The default recommendation.
Heavy
Complex multi-document reasoning
Only where justified.
Same input, same answer.
No sampling, no randomness. Determinism is a requirement, because the accuracy standard only means something if the system behaves the same way tomorrow.
Measured, not assumed.
Accuracy is recorded on your held-out cases at each size and precision, and the choice is written on the solution’s model card.
Adapted inside the perimeter.
Trained on a workstation in your building, signed, and moved to production through a one-way transfer. Nothing calls out.
Edge inference research
Measured honestly
Separately from REMIT, the edge-inference work behind FORGE and HELIX targets hybrid state-space architectures on Apple hardware, and is published with the numbers that did not work.
Featured release
Falcon-H1-7B-FORGE-v2
- Base
- tiiuae/Falcon-H1-7B-Instruct
- Architecture
- Hybrid Mamba-2 SSD + attention, 7.59B params, 44 blocks
- Compression
- 2.06 bits per weight average, mixed-precision PTQ
- Size
- 3.48 GB, from 15.18 GB at F16
- Runs on
- stock llama.cpp, unmodified
Design philosophy
We prioritise factual retention over bit purity. Uniform ternary quantization left the model fluent and confidently wrong — it named Sydney as the capital of Australia. The Hessian explained why: ssm_out, the projection writing the Mamba-2 scan output, carried an outlier flatness of 6.78 against roughly 0.33 everywhere else. Holding it and ffn_down at Q6_K costs 0.23 GB and returns the model's world knowledge.
Quality
| Build | Size | Wikitext2 ppl | Factual probes |
|---|---|---|---|
| F16 | 15.18 GB | 6.5875 | 12/12 |
| v1 — ssm_out ternary | 3.25 GB | 10.7590 | 9/12 |
| v2 — ssm_out at Q6_K | 3.48 GB | 9.1633 | 12/12 |
Two caveats we will not bury. The v2 run changed two variables at once — preserving ssm_out and raising calibration from 32 to 128 sequences — so the recovery cannot be attributed between them. And there is no same-size Q4_K_M baseline, so “better than conventional 4-bit at this footprint” is not a claim this data supports.
Performance
| M5 Max | iPhone 14 Pro Max (A16) | |
|---|---|---|
| Model load | ~2 s | 6.2 – 8.5 s |
| Prefill | 1941.8 tok/s @ 2048 | ~45 tok/s @ 1073 |
| Decode | 74.6 tok/s | 5.5 tok/s |
| Peak footprint | 3.73 GB RSS | 0.31 GB phys_footprint |
Running it
Requires a sampling chain — temperature 0.7, top_p 0.9, and repeat penalty 1.15 over a 2048-token window. The window, not the penalty strength, is the lever: at llama.cpp's default of 64 the sampler cannot see past the current turn, and cross-turn repetition measures 45% against 17% at 2048. It reduces looping rather than eliminating it.
On iOS the binding constraint is Metal's working set, not the jetsam limit — use n_ubatch = 256 on 6 GB devices. The open issues are listed in full on the model card.
The stack that produced it
Three open repositories, none of them a commercial product. The quantizer, the kernel, and the client we used to measure both.