01 Primary source
Moonshot AI
Kimi K3
A native multimodal 2.8T-class model for long-horizon coding, knowledge work, and million-token contexts.
TOTAL2.8T
ACTIVE104B
EXPERTS896
Long-horizon agentsView intelligence
DEPLOYMENT ASSURANCE / EXPERT-ROUTED AI
Prove it before
you deploy.
Plan an exact deployment. Run the open gates. Verify the resulting evidence. Share a portable receipt whose limits remain attached.
LOCAL DEMO Deterministic fixture bytes · no upload · no model-performance claim
15 sourced claims0 admitted controlled runs0 fabricated scores
Independent deployment intelligence from LockedIn Labs
PINNED ARTIFACT MAP REGISTRY FACTS / V1.0.0
RROUTER
E01
E02
E03
E04
E05
E06
E07
E08
8 OF 128 EXPERTS ACTIVE / TOKEN8 × H200 NODE VIEW1 GPU MIN. / CHECKPOINT ONLY
PINNED ARTIFACT61.1 GB
EXPERT ROUTING8 / 128
STATIC LOWER BOUND1 GPU MIN.
Qwen3-30B-A3B · sourced registry facts + calculated static fit · topology illustration, not measured runtime data
PLAN · EXACT ARTIFACTRUN · DEPLOYBENCH V0.1VERIFY · PASSPORT V0.2SHARE · PORTABLE RECEIPT
ONE SYSTEM / MULTIPLE ENTRY POINTS
PLATFORM V0.1
Enter wherever the decision starts.
The website is the visual workbench. The protocol, runner and developer surfaces make the same decision system portable.
01AVAILABLE NOW
Assurance Engine
Bind an exact artifact to hardware, runtime, workload and SLA. Calculate the static floor and open validation gates.
02RUNNER V0.1
DeployBench
Measure endpoint TTFT, request latency, success and aggregate throughput without retaining prompts or responses.
03START HERE · 90 SEC
Deployment Passport
Pack compatible trials under one digest, verify summaries and signatures locally, and keep every reproducibility gap attached.
04SOURCED CORPUS
Model intelligence
Inspect artifact-exact architecture and owner-reported evidence with every missing field kept visible.
05WORKSPACE PACKAGES
Developer stack
Use the same decision contract through the CLI, benchmark runner, REST API, TypeScript SDK and MCP server.
THE DEPLOYMENT GAP
01 / 05
37B active does not mean 37B resident.
Sparse activation reduces compute. It does not erase model weights, KV cache, runtime buffers, replication, or all-to-all traffic.
MOEModels separates can load, can run, can scale, and makes economic sense— because they are four different answers.
01
FACTS
Every model specification links back to primary evidence.
02
CALCULATIONS
Every derived value exposes its exact inputs and formula.
03
DECISIONS
Every result ends with a bottleneck and a next action.
MODEL INTELLIGENCE
02 / 05
Five artifact-exact
MoE records. One evidence layer.
Sourced architecture and hardware facts, with licensing, runtime compatibility and deployment unknowns kept visible.
Open model index02 Primary source
DeepSeek
DeepSeek V4 Pro
A frontier-scale model combining fine-grained sparse experts with a million-token context for agentic work.
TOTAL1.6T
ACTIVE49B
EXPERTS384
Frontier open reasoningView intelligence
03 Primary source
Z.ai
GLM-5.2
An open flagship designed for long engineering trajectories across a sourced million-token context.
TOTAL753B
ACTIVEUnknown
EXPERTS256
Long-horizon codingView intelligence
04 Primary source
Gemma 4 26B A4B IT
A compact multimodal MoE with 25.2B total and 3.8B active parameters, designed for accessible inference.
TOTAL25.2B
ACTIVE3.8B
EXPERTS128
Single-accelerator baselineView intelligence
Specifications shown above are sourced from pinned manifests, configurations, model cards, or technical releases. No benchmark ranking is implied.
MOE FIT CHECK
03 / 05
Turn an exact artifact into an honest lower bound.
Select a pinned checkpoint, hardware target, topology, and declared reserve. See what is proven—and what remains unknown.
Exact manifest bytes + integer math. No generated throughput, capacity, or price.
Deterministic registry engine
Will the checkpoint fit?
Test an exact, pinned artifact against advertised accelerator memory—then keep every unsupported conclusion visibly unknown.
Sourced + calculated
02 · Fit decision
Checkpoint cannot fit
Kimi K3 · 896 / top-16 experts · 1M context · 8 × NVIDIA H200 SXM · 13% reserve
Decision boundaryCheckpoint bytes only
01Baseline fails
Load
The checkpoint alone requires at least 13 GPUs at this reserve, before runtime allocations.
02Evidence required
Run
A baseline pass does not prove loader, kernel, quantization, sharding, or expert-parallel support.
03Unknown
Scale
No measured workload profile is attached, so throughput, latency, KV demand, and skew are not projected.
04Unknown
Economics
No dated provider, region, or utilization record is selected, so the engine emits no invented cost.
03 · Explainable topology
Requested
Artifact
1.6 TBtensors
Reserve
Experts
Compute8 × H200
- Checkpoint tensors
- 1.6 TB exact manifest metadata
- Usable / GPU
- 122.7 GB after declared reserve
- Device lower bound
- 13 before node rounding
- Node-rounded
- 16 GPUs 2 × 8-GPU nodes
RequestedThe topology you asked the engine to test against the checkpoint baseline.
View calculation anatomy
- Exact artifact tensor bytes
- 1,560,860,324,864
- Advertised memory / GPU
- 141 GB
- Declared reserve / GPU
- 18.3 GB
- Usable memory / GPU
- 122.7 GB
usable = floor(advertised bytes × (10,000 − reserve bps) / 10,000); minimum GPUs = ceil(checkpoint tensor bytes / usable bytes). A failure is conclusive for this no-offload baseline; a pass is only a candidate for runtime validation.
Artifact evidence: Kimi K3 pinned tensor manifest ↗
BENCHMARK EVIDENCE
04 / 05
A leaderboard should show its work.
Fifteen owner-reported claims keep model scope, settings, sources and missing context attached. No controlled run has crossed the comparison gate yet.
Enter the evidence labEVIDENCE CLASSMETHODSTATUS
Owner reports15 sourced claimsAVAILABLE
Controlled endpoint runs0 admitted packsOPEN
Verified Deployment Passports0 admitted packsOPEN
Independent reproductions0 reproduced packsFUTURE
Calibrated configuration corpus0 configurationsFUTURE
DeployBench measures. Passport verifies the receipt.Admission and reproduction remain separate.
OPEN INFRASTRUCTURE
05 / 05
One engine.
Every interface.
Workbench, CLI, runner, REST API, TypeScript SDK and MCP server share the registry, evidence classes and deterministic planning contract.
moemodels / fitREGISTRY V1
$ npm run moemodels -- fit \
moonshotai/Kimi-K3 \
nvidia/h200-sxm-141gb \
--gpus 8 --reserve-pct 13
CHECKPOINTFAIL × 8
TOPOLOGY13 → 16
RUNTIMEUNKNOWN
Real offline CLI · same registry and integer engine as the hosted Fit Check
RESEARCH DESK
Read past the parameter count.
FIELD GUIDE 01
Active parameters are not a deployment plan
A practical guide to resident weights, KV cache, runtime overhead, and the four different meanings of “fits.”
SYSTEMS NOTE 07
Expert parallelism, without the hand-waving
How topology, all-to-all traffic, load imbalance, and hot experts reshape the economics of sparse inference.
BUYER BRIEF 03
Cloud endpoint or private cluster?
A decision framework for comparing utilization, privacy, operational burden, and three-year total cost.
START WITH THE PROOF
Break the receipt.
Watch trust fail closed.
Verify a deterministic Passport fixture, change exactly one byte, and see the content address reject it—all locally in your browser.
Run the 90-second check