AI models that run where your data lives.

Hypereum designs and operates specialised AI models: a proprietary fleet powering Hivemind AX, and open releases anyone can verify. The first family is compliance. Every model is built to run in your jurisdiction, on your hardware.

The models that power Hivemind AX for UK/EU regulated work: a proprietary fleet, and one open release anyone can verify.

The fleet, by role.

Planner

Reads a mission brief and decomposes it into a dependency-ordered task plan. Each task carries its scope, the evidence it needs, and where it sits in the execution order. Executors pick the plan up from there.

mission in · dependency-ordered task plan out

PROPRIETARY RELEASE-GATED

Executors

Perform the compliance analysis task by task, against the supplied evidence. Each task returns findings with severity, remediation effort, and the regulatory references they rest on. Raw executor output flows to the synthesizer, never straight to the report.

task + evidence in · findings, severity, regulatory refs out

PROPRIETARY RELEASE-GATED

Synthesizer

Takes every executor finding for a mission and aggregates it into mission-level risk. It resolves overlaps between tasks and weighs severity across the whole mission. What leaves this stage is the view a reviewer reads.

executor output in · mission-level risk out

PROPRIETARY RELEASE-GATED

Evaluator

Scores the quality of what the pipeline produces, finding by finding. Every score lands on a fixed schema, dimension by dimension, before anything is accepted. This is the one role we ship in the open: HivemindEval.

finding in · six-dimension scores out

OPEN · APACHE-2.0 See the benchmark

Roles are stable. The models behind them are versioned, and rotate only through the gates below.

Five gates. Every release.

No fleet model reaches production without passing the acceptance gates published for HivemindEval: score variance across graded inputs, spread between quality tiers, strict tier ordering, and jurisdiction behavior checks.

G1 Variance overall scores must vary across graded inputs; a constant scorer fails
G2 Per-dimension variance every scored dimension must vary, not just the overall number
G3 Spread a clear gap between the best and worst quality tiers
G4 Monotone ordering average scores must rank the tiers in strict order, with zero violations
G5 Jurisdiction sensitivity findings that cite the wrong jurisdiction must score low

The gates are published in full on the model card

Every number on this site follows one rule: it goes public only when anyone can recompute it. HivemindEval's results qualify, because the weights, the benchmark subset, every model's raw predictions and the scoring code are all open. Our proprietary fleet models pass the same internal gates before release, but their results descend from a validation apparatus we still label provisional, and there is no public benchmark on which you could rerun them. Publishing those numbers would be marketing, not measurement. The same standard will govern anything we publish in the future.

HivemindEval: the open proof point.

PROVISIONAL · pending independent ≥3-expert validation

Our evaluator is open-weight. So is the benchmark. So are every model's raw predictions. Check us.

Format validity, before judgment quality.

Base model, native prompt 0/212 · prose, not JSON
HivemindEval v2.1 211/212 · valid schema

Same hardware. Same prompt class. Format reliability is what the fine-tuning buys, reported separately from judgment quality, on purpose.

Five models, one benchmark.

Qwen2.5-72B-Instruct-AWQ (rubric) 0.985
statistical tie · overlapping 95% CIs
HivemindEval v2.1 (native) 0.984
Qwen3-8B (rubric) 0.973
Llama-3.3-70B (rubric) 0.965
Qwen3-8B (native prompt, no fine-tune)
HivemindEval benchmark: Full-212 board, all four metrics
Model Tier concordance (95% CI) Kendall τ Defect separation Schema-valid
Qwen2.5-72B-Instruct-AWQ (rubric) 0.985 (95% CI [0.977–0.993]) 0.801 0.602 212/212
HivemindEval v2.1 (native) 0.984 (95% CI [0.973–0.993]) 0.817 0.591 211/212
Qwen3-8B (rubric) 0.973 (95% CI [0.959–0.985]) 0.751 0.709 208/212
Llama-3.3-70B (rubric) 0.965 (95% CI [0.950–0.980]) 0.764 0.616 173/212
Qwen3-8B (native prompt, no fine-tune) 0/212
HivemindEval benchmark: Core-68 board, all four metrics
Model Tier concordance (95% CI) Kendall τ Defect separation Schema-valid
Qwen3-8B (rubric) 0.991 (95% CI [0.977–1.000]) 0.788 0.639 67/68
Qwen2.5-72B-AWQ (rubric) 0.990 (95% CI [0.980–1.000]) 0.833 0.549 68/68
HivemindEval v2.1 (native) 0.979 (95% CI [0.958–0.998]) 0.843 0.615 67/68
claude-sonnet-5 (rubric) 0.976 (95% CI [0.951–0.996]) 0.777 0.562 66/68
Llama-3.3-70B (rubric) 0.964 (95% CI [0.932–0.988]) 0.798 0.667 56/68

Check it yourself.

FULL SET 4f2dbff7… withheld · committed by hash
PUBLIC SUBSET b51e721e…

$ python3 eval/score_benchmark.py

recomputes every CORE-68 cell, bit for bit, seed 7, no GPU

Every CORE-68 number is recomputable offline from published data. The full 212-item set is withheld and committed by hash above.

Sovereign deployment.

Frontier-class verifier output. Data that never leaves your jurisdiction. Hardware you own.

24 GB
GPU
single-stream, minimal headroom
~40
tok/s
single-stream, observed
~23.4
GB VRAM
observed, vLLM at --max-model-len 8192

24 GB deployment is single-stream with minimal headroom; use a larger card for concurrency.

Verifier quality is provisional; see the benchmark above .

Get in touch