REPRESENT · RECORDED RUNS · KIMI-LINEAR-48B-A3B
DON'T CHOOSE A QUANTIZATION. DECLARE THE BEHAVIOUR YOU WANT TO PRESERVE.
Quantization asks you to pick a format and hope. REPRESENT inverts it: freeze a behavioural contract, let a composition-aware search discover the cheapest physical topology that provably satisfies it — and verify the composed model, not tensor by tensor.
What is REPRESENT?
REPRESENT is VINDEX3's optimizer: it searches over candidate physical representations (BF16, Q8_0, Q6_K, Q4-class…) for each region of a model, subject to a frozen behavioural contract, and returns the cheapest topology whose COMPOSED behaviour passes at authority scale. Quantization is one of the mechanisms it uses; the decision-making — which representation, where, in which combination — is the point.
How is REPRESENT different from quantization?
Quantization converts a model into one format and measures afterwards. REPRESENT declares what must be preserved first — a contract over KL divergence, displaced probability mass, and routing consequence at a fixed evidence scale — then treats every format as a candidate and every region independently. The output is not a converted file but an earned topology: BF16 where behaviour demands it, cheaper encodings where evidence permits them, verified as a whole model.
The search is real, and its first earned topology is three layers wide.
THE RECORDED RESULT — STRICT
Kimi Linear 48B under the frozen kimi-logit-v3 contract. Uniform Q6 at the probed depths: refused. The discovered strict topology keeps layers 0–23 at BF16 and takes layers 24–26 to Q8_0 — earned at 8,192 teacher-forced positions with KL p99 4.153e-4 against a 1e-3 limit, removing 5.1 GB, decoding 2.1–2.8% faster beside its own in-session BF16 baseline. Small on purpose: strict is the fidelity reference profile, and every layer it excludes was excluded by measurement.
kimi-linear-48b · routed-expert precision by layer · click a layer
LAYER 25 · ROUTED EXPERTS · Q8_0 in the strict topology
BF16
canonical — the yardstick both arms are measured against
Q8_0 · KL p99 vs limit 1.000e-3
1.446e-4 · PASS · 8192 positions
Q6_K
1.078e-3 · REFUSE · 8192 positions
KDA projections at Q8_0 (second family, same contract): 8.837e-5 — PASS — cheaper than the expert cell at the same depth.
Q6_K independently exceeded the behavioural budget here. Q8_0 earned the slot at authority scale.
Individually safe ≠ safe together.
Why can individually safe quantizations fail when combined?
Because a model is not a bag of independent tensors. Each substitution displaces the hidden state; later routing decisions read that displaced state; errors interact through the model's own control flow. Measured on Kimi Linear: three composed maps whose every member passed alone all failed the contract — one super-additively, its composed displacement 1.9× the sum of its members. A per-layer sensitivity table cannot see this. Only measuring the composed model can.
FOUR COMPOSED MAPS · ONE FROZEN CONTRACT · RECORDED
The composition factor is not a constant: measured ratios of composed to summed displacement ran 0.52, 0.55, 0.77, 0.88 — and 1.90, super-additive, on the third map above. No scalar correction ranks maps. The interaction is structural: a substitution's error cascades through every later routing decision, so the same layer costs ~6× more inside the wrong company, and the last routed layer composes nearly free because nothing downstream can amplify it.
WHY THIS IS NOT A SENSITIVITY SPREADSHEET
The layer the search left out is the proof. L26's cheapest individually-passing representation was Q6_K — and Q6_K at L26 belongs to no admissible composed map, because it consumes most of the whole behavioural budget by itself. Q8_0 there is seven times cheaper behaviourally and composes nearly free, since no later routed layer exists to amplify it. A recipe built from isolated measurements would have shipped the wrong bytes.
What is a behavioural fidelity contract?
A frozen, versioned gate the candidate must pass — not a vibe, a program. kimi-logit-v3 requires at least 4,096 teacher-forced positions and bounds KL p99, covered probability mass, displaced top-1 and top-10 mass, and routed-mixture movement. It judges consequence, not counts: at authority scale a passing map showed 1,041 top-10 reorderings, almost all near-ties — a count-based gate would have refused a good map, and diagnostic-scale counts were measured too noisy to certify anything. The contract is the authority; the search only proposes.
Representation topology isn't limited to experts.
Does REPRESENT only quantize expert weights?
No — and this is where the decode economics live. On Kimi Linear, routed experts are 47.4% of decode bytes; the KDA recurrent projections are another 25.4%. Both families are now opened: KDA projections at Q8_0 measured CHEAPER behaviourally than expert weights at the same depth (KL p99 8.8e-5 vs 2.3e-4), because their error enters through a smooth recurrence instead of a discrete routing boundary — and the first cross-family map, experts plus KDA together, passed the full contract at 8,192 positions with the two families' errors composing sub-additively.
DECODE BYTES PER TOKEN · KIMI-LINEAR-48B · BF16 · 6,209 MB TOTAL
opened — strict map earned at 8,192 positions
opened — cross-family map earned at 8,192 positions
next
next
next
—
This ledger is calibrated, not hypothetical: the earned map removes ~3.7% of decode traffic and measured a 3.2–3.7% GPU-time reduction beside its own in-session baseline — bytes convert to time at the predicted rate, across both families opened so far. The route to large decode gains runs through every row of this table at once. REPRESENT optimizes all of them together.
EVIDENCE
The contract catches what diagnostics cannot — RECORDED · 8,192 POSITIONS
A balanced-bracket candidate that looked benign at 256 positions (two near-tie argmax flips) produced 63 argmax flips at authority scale, the worst surrendering 18% of its answer's probability. Small-sample consequence counts are noise; the contract's evidence-scale floor is what makes 'preserved behaviour' a checkable claim rather than a hope.
Depth is a structural signal, in both families — RECORDED
Expert Q8_0 passes from L20 back and collapses toward the front. KDA Q8_0 is nearly free in the late tower, marginal at L16, and refuses by L13 — where a new mechanism appears: displacement accumulating across token distance through the recurrent state. Same contract, different physics per family; the search inherits both maps.
The byte ledger is predictive, not aspirational — RECORDED · IN-SESSION BENCH
The earned cross-family map removes ~3.7% of decode traffic and measured 1.036–1.043× decode beside its own same-session BF16 baseline, GPU time 27.2–27.4 → 26.4–26.5 ms/token — the reduction the ledger predicted. Four sessions, two power states, reproduced after a full container re-encode (the strict authority run reproduced to the last digit).
Search once. Keep the result forever.
THE SHAPE OF THE SURFACE — A WORKED SHAPE, NOT A RECORDED RUN
The CLI surface this chapter is heading toward: `vindex represent search model.vindex3 --contract fidelity --objective min-bytes` returning the earned topology with its composed verdict, then `vindex represent save kimi.represent` persisting identity, topology, evidence and benchmark as one durable object. Today the search runs as instrumented harnesses and the artifact is assembled by hand; the object it produces — source identity, contract hash, evidence-bank identity, per-cell verdicts, composition history, benchmark — already exists for the Kimi result. The verbs are being productized in that order.
ROADMAP — ONE CANONICAL MODEL, MANY EVIDENCE-DERIVED FORMS
Design direction, not shipped: a REPRESENT result is small enough to publish beside the canonical checkpoint — topology, contract, evidence — with derived physical banks cached content-addressed and shared between profiles that select the same bytes. One canonical model on the Hub; many evidence-derived physical representations materialized on demand; no zoo of hand-named quant files. The exported forms can include stock containers: a discovered per-layer map is expressible as a GGUF recipe, so the optimizer's answer travels beyond this runtime.
Can a REPRESENT result be saved and reused?
That is the design's centre of gravity: the search is expensive once and amortized forever. The result object carries the source identity, the frozen contract, the evidence-bank identity, every per-cell verdict, the composition history and the benchmark — and it was re-earned unchanged after the canonical container was re-encoded, which is the property that makes it durable model knowledge rather than an experiment log. The packaged .represent artifact and its registry distribution are roadmap; the object's contents are already real.
Can REPRESENT export to GGUF?
Per-layer and per-projection topologies are expressible in stock GGUF, whose tensors each carry their own type — so a discovered map at that granularity can compile down to a recipe that runs in llama.cpp unmodified. Finer granularity (per-expert, sub-tensor, or representation-times-residency plans) exceeds what stock containers can carry and stays native. Export is roadmap; the granularity analysis is done.
TRY IT — THE NEIGHBOURING FACTS, IN THE EXPLORER
Quantization is a mechanism. REPRESENT is the optimizer.
The map the quantization chapter could store and the discovery chapter could not write — this is what writes it: a contract that judges, a search that composes, and evidence that outlives the container it was measured in.
PUBLISHED 31 AUG 2026 · VERSION 3.0 Candidate
CITEsite build 43a1b6b · built 2026-09-02
CITE THIS
Web page · 3.0 Candidate
Hay, C. (2026). REPRESENT. In VINDEX3 Specification (Version 3.0 Candidate). VINDEX3. https://vindex3.org/represent
A chapter of the specification, citable on its own. Cite the version rather than the site — the canonical reference, and how to use it, live at /cite.