Moonshot AI · exact checkpoint record

moonshotai/Kimi-K3

The named repository revision is the boundary for every fit statement. Similarly named checkpoints, unofficial conversions and unrecorded runtime configurations require separate testing.

Review system evidence Back to model centre

A private deployment starts with the permitted data path, access policy and logging boundary.

Reviewed 28 July 2026 Review by 28 August 2026 0 measured system matches

Checkpoint identity

Checkpoint specification

Provider and family: Moonshot AI · Kimi K3
Checkpoint: moonshotai/Kimi-K3
Repository revision: Repository state retrieved 2026-07-28; pin an immutable commit before deployment
Input / output: text, image → text
Total / active parameters: 2.8 trillion total parameters; 104 billion active parameters
Native numerical format: Native MXFP4 weights and MXFP8 activations

Dated artefact inventory

1,560,936,091,448 bytes across 96 safetensors files

Dated repository-file total, not a prediction of loaded memory, KV-cache headroom or runtime overhead.

Documented context ceiling

Up to 1,048,576 tokens in the model specification

A documented model limit is not proof that a GPU RIGS configuration can serve this context.

System-by-system register

What is known across the eleven-system range

Measured support is shown separately from documented and calculated candidates. A capacity calculation alone does not establish compatibility.

System	Class	GPU memory	Evidence state	What the record proves
Team 32	SME workstations	32GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Company 64	SME workstations	64GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Studio 96	SME workstations	96GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Studio 192	SME workstations	192GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Value Rack 128	Enterprise PCIe racks	128GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Value Rack 256	Enterprise PCIe racks	256GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Enterprise 384	Enterprise PCIe racks	384GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Enterprise 768	Enterprise PCIe racks	768GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
H200 1.1TB	Enterprise PCIe racks	1,128GB aggregate	Not evaluated	No checkpoint-to-system fit test is recorded. This is neither a support promise nor a rejection.
Frontier Native 2.3TB	Frontier partner platforms	2,304GB aggregate	Calculated capacity candidate	HGX B300's documented 2.30TB aggregate HBM3e exceeds the dated 1.5609TB repository weight total. This arithmetic establishes only a memory-capacity candidate, not a successful load, usable context, throughput, latency or concurrency.
Frontier Rack 20TB	Frontier partner platforms	20,736GB aggregate	Calculated capacity candidate	GB300 NVL72's documented 20TB aggregate GPU memory is a capacity candidate for this checkpoint. Its commercial purpose is rack-scale throughput and concurrency, neither of which is measured here; checkpoint fit alone does not justify the platform.

Aggregate GPU memory is an inventory value. It is not automatically a single pooled-memory address space, and it does not prove a successful model load.

Runtime evidence

No reproducible runtime benchmark is available.

Engine version, container digest, driver, CUDA version, parallelism, context, batch, concurrency and benchmark logs must be captured together before performance or measured support can be confirmed.

Primary source records

Source state was retrieved on 28 July 2026 and is scheduled for recheck.

Diagram showing an approved request, a local service, an approved store and a policy-controlled data path — A private deployment starts with the permitted data path, access policy and logging boundary. GPU Servers technical illustration.

GPU server remote management dashboard with system status, access logs and sensor monitoring panels — Supplier screenshot of the platform management interface. The final management features and access policy depend on the ordered system. OEM supplier reference image.

Diagram showing approved documents moving through a searchable index to an answer with a source citation — A retrieval workflow should connect each useful answer to approved source material and defined refusal behaviour. GPU Servers technical illustration.

Diagram of an evidence pack containing an asset schedule, burn-in record, health readings, workload test and admin guide — A credible handover records the supplied assets, checks, operating evidence, instructions and unresolved items. GPU Servers technical illustration.