Nine-channel interface — PROMOTED
The physical Hyphae channel adapter passed every typed gate at 100%, preserved the frozen parent exactly (491/491 outputs), and was promoted as a governed component — with mandatory deterministic authority override.
Model bodies
MARS is not a checkpoint
MARS is the identity; model bodies are replaceable artifacts, pinned by SHA-256 and admitted through gates. Two bodies matter today: MARS-1 running now, and MARS-T being distilled for CPU.
Running today
MARS-1 — quantized for consumer hardware
The first MARS body is a fine-tuned Gemma-family model. Its Q4 quantization passed the consumer gate on an 8 GB laptop GPU — measured, not projected.
RTX 5050 gate — measured
Measured- Reference GPU
- NVIDIA GeForce RTX 5050 Laptop · 8 GB
- Model artifact
- 5.29 GB (Q4_K_M)
- Prompt processing
- 1,616 tok/s
- Generation
- 50.2 tok/s
- Peak VRAM used
- 3,098 MiB
- VRAM budget
- 8,151 MiB
sha256 · f56cb110f2b6b0f4…
What passed — and what did not
Promotion is earned per capability, never assumed. Two decisions define MARS-1 today:
Q4 as complete body — REJECTED
The Q4 parent scored 86.1% on source-pointer selection against a ≥99% gate. Decision: keep Q4 as an experimental local parent, remediate — do not weaken the gate. This rejection is published like every pass.
The destination
MARS-T — ternary weights, CPU-native
The future body uses three-valued weights {-1, 0, +1}: multiplication-free inference that runs on CPU and RAM — no GPU required. Distillation is preregistered and gated; only verified behavior gets distilled.
The contract
- Weights restricted to {-1, 0, +1} — additions instead of multiplications.
- Teacher trajectories are captured only from verified, policy-authorized runs.
- No mutable factual corpus is ever distilled into weights.
- Promotion requires exact action/abstention gains with confidence intervals — and rollback.
Receipt · mars-t-bounded-student
-1 0 +1
Current status: a bounded ternary policy slice is admitted (action selection + evidence-absent abstention, EN/ES). It is explicitly not the full MARS-T model — and the receipt says so.
Reversible by construction
Every adaptation can be rolled back
Adapters are trained beside frozen parents, never inside them. Unload the adapter and the exact original model returns — verified by digest. Zero parent gradients is a measured gate, not a slogan.
Intelligence you can audit.
MARS, HiveMind and Hyphae are being built in the open by Celiums Solutions — local-first, fail-closed, and receipted end to end.