jev-nano

The nano tier of local Jev. A /v1/systemone-compatible typed-decision server — noul / choice / score questions with probabilities — running on a 7M-parameter model, CPU-only, in an 8MB tarball. No GPU, no Python, no dependencies beyond a Linux x86_64 box.

Written entirely in MFL (machin) — the model runtime, the server, and the decision heads compile to a single static native binary through machin's C backend. There is no Python anywhere in the inference path.

Author: Javier Leandro Arancibia

Point any Jev/TypeSafe SDK client at http://127.0.0.1:8097 via base_url.

Install & run

curl -fsSL https://raw.githubusercontent.com/javimosch/jev-nano/main/install.sh | bash
./demo.sh   # smoke test

or manually:

curl -fsSLo appliance.tar.gz \
  https://github.com/javimosch/mtlm-router/releases/download/v0.3.0/mtlm-router-v0.3.0-linux-amd64.tar.gz
tar xzf appliance.tar.gz && cd mtlm-router-v0.3.0-linux-amd64 && ./start.sh

Use

curl -s http://127.0.0.1:8097/v1/systemone \
  -H 'content-type: application/json' \
  -d '{"state":"Payouts failing for 3 days, help ASAP.","model":"jev-latest",
       "questions":{"u":{"type":"noul","instructions":"Does this convey urgency?"}}}'
# -> {"answers":{"u":{"type":"noul","noul":0.989}}, ...}

Real Jev's published answer to that exact question: 0.95. Ours: 0.99.

How it compares

	jev-nano	jev-local (OSS)
weights	7M params	1.5B–9B instruct models
download	~8MB	6–18GB
hardware	any CPU	GPU box or 16–24GB RAM
scoring	head + candidate logprob	per-candidate logprob
abstention	calibrated conformal gate (set_size, qhat, error budgets)	max-prob confidence
contract	/v1/systemone drop-in — 19/19 contract checks (request shapes, validation, error types)	/v1/systemone

jev-local's own eval found every model under ~1.5B collapses on choice/score questions. jev-nano is the tier below that floor — a frozen trunk plus small trained heads instead of raw logprob guessing.

Contract conformance

bench/systemone/contract.sh in mtlm-router exercises the wire contract: string/object/array state, criteria-map and options-array choice forms, score legends, multi-question answers, every documented validation error, and max_tokens_exceeded on context overflow. 19/19 against v0.3.0.

Extra endpoints

Same binary also serves /v1/route (dispatcher with abstention/delegation), /v1/decide /v1/noul /v1/score /v1/assess (trained head lanes), /v1/chat/completions, /v1/feedback, /guide, /llms.txt.

Honest benchmark (JevBench public set)

Ran JevBench's 231 public items through this server (harness in bench/): 28% on the easy+original tiers (120 items — the hard tier's ~3.7k-token states exceed this model's 384-token context). Raw candidate-logprob scoring on a 7M-parameter trunk is real but shallow: it agrees with hosted Jev on clear noul/urgency calls and is near chance on fine-grained multi-way choices. Compare local-jev's 80.5% on the full 231 using a 1.5B+ instruct model — that is the honest price of 8MB.

The accurate path at this size is the trained-head lanes (/v1/route, /v1/decide): a head fitted on your labels reaches 90-98% on its taxonomy at the same 7M/CPU footprint — that is the product.

What this is

The open demo of mtlm-router — a self-hosted typed decision layer for production routing: per-client heads, calibrated abstention (conformal qhat, selective-risk bounds), live recalibration via /_qhat, decision logs, feedback loop. Repo: javimosch/mtlm-router · runtime: javimosch/machin-anvil · writeup: how the conformal gate works.

Hosted demo API: https://api.mtlm-router.intrane.fr (metered, free starter credit).