Inspect before allocating.
Read actual GPU occupancy, available RAM and tensor headers. Review the exact device, model revision, local image, context, RAM cap and deadline. Existing services stay untouched.
MiMo-V2.6 / ONE-CARD EXPERIMENT
One GPU.
The full
checkpoint.
Keep the model and the benchmark.
Change where the weights live.
An inspectable test for one card in your Blackwell tower, with host-memory offload, an explicit resource budget, and your own scorer.
Local inspection first. Model loading needs your approval.
RESIDENCY, NOT MODEL REDUCTIONSCHEMATIC
ONE SELECTED DEVICE01
GPU96 GB reference card
RTX PRO 6000BLACKWELL
CPU offload · transfer cost measured
HOST RAMCapacity and headroom checked locally
COMPLETE LOCAL CHECKPOINTOTHER GPU EXCLUDED
THE ACTUAL QUESTION
Can one GPU plus sufficient host RAM run this complete checkpoint under your benchmark’s real constraints? The runner is ready to test that. Native MiMo inference: not yet run.
01 / PREPARE
Your machine. Your test.
Choose the model and evaluation path.
Copy one brief into your existing local agent.
1.02TTOTAL PARAMETERS
42BACTIVE PER TOKEN
First establish real load and generation behavior. A canary produces no benchmark score.
USES YOUR EXISTING
Linux or WSL2 · Python 3.10+ · Docker + NVIDIA runtime · complete local checkpoint · compatible local vLLM image
LOCAL AGENT BRIEF
No weights, packages or images downloaded automatically.
Download complete kit ↓
Inspect → review the plan → approve a canary → benchmark.
02 / EXECUTION
A small footprint.
A complete accounting.
Model custody, execution and scoring remain separate. A refused load is a useful result.
READ-ONLY FIRST
Give one load a bounded budget.
After approval, hash the complete checkpoint and test the real loader. Exactly one GPU is exposed to the isolated container. The runner retains errors and cleans up only its own job.
Keep your evaluation intact.
Use frozen independent requests or your existing harness through a temporary local endpoint. Preserve its scorer, tools, prompt formatting, reasoning budget and deadlines. Score with the original evaluator.
THE MEMORY BOUNDARY
Active parameters are not
the storage requirement.
Pro reports 1.02T parameters with 42B active per token. Routing can select different experts at different positions. This test retains the complete checkpoint and moves residency into real host RAM instead of deleting experts.
The current candidate uses vLLM CPU offload. It is not an NVMe-streaming port of the Kimi runner. Storage precision, loaded precision, KV state and transfer costs all affect fit. A missing RAM budget stays a blocker.
03 / INSPECT
Every moving part, open.
Read the code here before running anything.
Source previews fetch only this site’s static files.
v1/README.mdOpen raw ↗
Select a file to inspect its source. No local machine access is required.
Static source only. Selecting a file does not execute it.Get all files ↓
04 / EVIDENCE
The claim stops here.
Control tests qualify the controller.
The model still has to earn its result.
Runner control tests38 / 38 PASS
Synthetic checkpoint and mocked-process tests. No model inference.
Handoff + publicationBROWSER-TESTED
Copy, selection, source inspection, download and mobile checks. Inspect runs ↗
Native MiMo, one GPUNOT YET RUN
Load, memory use, throughput, numerical agreement and benchmark scores remain unmeasured.
What this carries forward from Aperture and Kimi
Aperture contributes permissioned hardware inspection and explicit placement. The Kimi experiments contribute source-bound state checks and accounting for constrained execution. Their Kimi-specific streaming kernels have not been ported to MiMo here.
The retained K3 strict-state result covered 93 layer caches at two accepted positions and reported 1.38× canonical improvement against a 588-second/token baseline. That is a narrow state-fidelity result, not a MiMo performance forecast.
What we know about the reference machine
Osmantic’s public lab lists a Blackwell tower with two 96 GB RTX PRO 6000 cards and vLLM. This candidate selects only one. The lab does not establish the host-RAM budget or current occupancy, so the local inspection supplies those facts before a load.
READY TO INSPECT