Explore releases Cavumo is being built to turn compatible models, private knowledge, and approved memory into capabilities your people, agents, and applications can use—and improve under your control.
Pre-release · No public downloads yetExplore working MVP features and the evidence behind them. “Coming soon” refers to public availability—not a released product or a delivery date.
Native MVP walkthroughs completed. Local workflows run in the development app.
RELEASE BOUNDARYStandalone installer, final first-use validation, signing and notarization remain pending.
Explore the system46.43% lower overall p95 latency versus Cavumo’s fixed configuration in the frozen Gemma/M2 study; identical outputs in 1,020 pairs.
RELEASE BOUNDARYScoped development benchmark, not a frontier-model speed comparison. Final production artifact needs retesting.
Read the benchmarkAn external local client can call the Engine; key revocation and transport constraints have been tested.
RELEASE BOUNDARYDeveloper tooling exists; public package, polished installation and a complete management CLI remain pending.
Explore developer interfacesLocal SDK transport and TypeScript integration checks pass against the development interfaces.
RELEASE BOUNDARYNo published packages or customer workload performance benchmark yet.
Explore developer interfacesApproved-memory persistence and unauthorized-agent isolation are covered by local workflow tests.
RELEASE BOUNDARYBusiness time savings are not measured; broader deployment validation remains pending.
Read the architectureTwo registered test agents share approved context with provenance; unauthorized retrieval and revocation are tested.
RELEASE BOUNDARYLocal service only today. The customer-owned multi-user server is a later phase.
See the sharing modelActual fitting and paired evaluation completed. A four-example test fixture measured loss 13.0833 → approximately 1.9292.
RELEASE BOUNDARYA tiny functional loss study, not answer accuracy. Compatible inference, owner activation and rollback are unfinished.
Read the fitting studyA clause candidate was trained and replicated; its ONNX artifact matched across exports. Strong-baseline comparison informed a non-promotion decision.
RELEASE BOUNDARYThe candidate did not pass its token-efficiency gate and remains inactive. No validated specialist release.
Read the experimentThe operational literal-citation baseline is integrated into the local Specialist console.
RELEASE BOUNDARYThe learned semantic-support model is not trained. Quote matching is not a truth or entailment guarantee.
Explore Verify researchRule-based guidance is operational locally, separate from grants and activation authority.
RELEASE BOUNDARYThe learned experiment-outcome predictor is not trained. No predictive benchmark is claimed.
Explore Adapt researchFeatures describe the local development MVP. Workflow checks are functional evidence, not performance benchmarks. Cavumo is proprietary software; compatible base models retain their own licenses. Commercial pricing and public release dates are not announced.
Cavumo’s services give agents continuity, shared understanding, and a controlled path to improvement. The same Engine enforces their boundaries.
Keep useful preferences and experience across work. Agents propose; people approve durable memories. Scope, source, retention, and training permission remain explicit.
A shared, versioned understanding for authorized agents. Approved project decisions, policies, and terminology stay connected to their owners and sources.
Fit a separate Gemma MLX adapter locally under a reviewed dataset and bounded grant. Compare held-out loss with the base model, cancel jobs, and recover interruptions. Behavioral approval, activation, and rollback are still in development.
Choosing a model is the beginning. The Cavumo Engine is the proprietary system being built to turn a compatible checkpoint into a governed, usable local capability.
Identity, scope, and purpose are checked before context is selected. Reading, sharing, and training stay separate.
Profiles become resource-bounded execution plans with token budgets, scheduling, pinned dependencies, and cache invalidation.
Track the model, context revisions, policy, adapter and evaluation evidence. Adapter activation and rollback remain under development.
01Context selection
02Local generation
03Evidence assessment
We’re developing compact proprietary models to help the wider system choose context, assess evidence, and recommend improvements.
Select relevant knowledge and approved memories within a context budget.
Learned scoring over local embeddings and retrieval signals; compare with hybrid search and reranking.
Target: at least 20% fewer input tokens, with no more than 2 percentage points of quality loss and 5% p95 latency regression.
Generation 1 · Context has a trained, inactive candidate. Verify and Adapt have operational rule baselines; learned models are not trained yet. No specialist is validated for release.
A recurring failure should lead to a tested decision. The current experiment service fits Gemma adapters and evaluates held-out loss. We are building the remaining behavioral checks and owner-approved activation lifecycle.
Explore the research agendaTry context or configuration changes when they address the failure. Cavumo Adapt 1 is a proposed advisory model for choosing the next experiment.
The MVP compares base and adapter loss on held-out examples. Dataset, model, grant, and evaluation records stay connected; original checkpoints are preserved.
Training requires a bounded grant and compatible hardware. Candidates remain inactive today. Compatible inference, behavioral regression rejection, owner activation, and rollback are release requirements still under development.
Choose your model, connect knowledge, and manage memory and agents in one considered desktop workspace. Designed for local operation.
Explore Cavumo StudioPre-release. Public downloads are not available yet.Proprietary software licensed for local use. Compatible base models keep their own licenses. Public installers and specialist weights follow compatibility, quality, security, and distribution gates.
Apple Silicon macOS
Coming soon · Public releaseCustomer-controlled local execution
Coming soon · Public releaseLocal specialist-model families
Coming soon · Public releaseTerminal · TypeScript · Python
Coming soon · Public releaseCommercial terms and offline entitlement details will be published before release. Performance targets are distinct from the scoped internal compiler study; learned specialist release validation is pending.
Read the evaluation gates
CAVUMO ENGINE
