Key takeaways
- Moonshot’s official Hugging Face repository was created on November 2, 2026 and now exposes Kimi K3.1 Pro weights, configuration and inference code. This is an artifact release, not merely a promise that weights are coming later.
- Moonshot describes a natively multimodal mixture-of-experts model with 2.8T total parameters, 104B active parameters and a 1,048,576-token context window. The repository exposes 96 safetensor shards, so self-hosting is a serious infrastructure project.
- The managed API IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Pro costs $2 input and $10 output per million tokens; Highspeed costs $4 and $20. Cached input and cache writes have separate rates.
- The custom Kimi K3.1-Pro License grants broad rights but is not a standard permissive license: qualifying model-as-a-service businesses need a separate agreement, and very large commercial products can trigger a display requirement.
- Trial the managed Pro route first unless weight control is itself a requirement. Treat Moonshot’s coding, 3D and finance benchmark results as vendor-reported until reproduced on representative workloads and infrastructure.
The weights are public, not “coming soon”
Moonshot AI has released Kimi K3.1 Pro as both downloadable artifacts and managed products. The official Hugging Face registry records the repository’s creation at 04:13:51 UTC on November 2, 2026. It includes 96 safetensor shards, configuration, tokenizer and processor assets, custom model code, a model card and a license. Moonshot’s launch page carries an October 24 dateModified value and had said full weights would arrive by November 4; the registry is the stronger evidence that the artifacts are now present.
The same model is available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Enterprise access is off by default for administrators, according to the announcement. AccessAllGPT verified public documentation and repository presence, not authenticated product availability or successful inference.
K3.1 Pro is enormous even with sparse activation
Moonshot describes K3.1 Pro as a natively multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion activated per token. Its 93-layer text stack combines Kimi Delta Attention with gated sparse attention, while MoonViT-V2 supplies the image encoder. The model card lists text and image input, MXFP4 weights, MXFP8 activations and a 1,048,576-token context window.
Sparse activation reduces per-token compute relative to a dense 2.8T model, but it does not erase weight storage, memory movement, interconnect or serving complexity. The public repository is proof of availability, not proof that a typical single-node setup can run it economically. Moonshot does not provide a buyer-neutral total-cost comparison against its API.
The API offers two speed and price tiers
The exact managed IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Standard Pro is $2 per million input tokens and $10 per million output tokens. Highspeed, which Moonshot claims is two to three times faster, is $4 input and $20 output. Both list cached input at $0.20 per million tokens; cache writes vary by model and five-minute versus one-hour TTL. Standard Pro supports batch inference at 60% of the standard price.
API access requires a successful top-up of at least $1. K3.1 Pro always uses thinking mode and accepts low, high or max reasoning effort, with max as the default. The guide says sampling values are fixed, public image URLs are unsupported, and web search is being updated and is not currently recommended for production. Those contract details belong in a migration test, not in a footnote.
The benchmark table is vendor evidence
Moonshot reports 65.7 on Terminal-Bench 4.0 and 92.7 on BenchCAD with tools, alongside results for several proprietary and open models. It also reports improvements on its internal long-horizon coding, finance and factual-accuracy evaluations. Unless otherwise specified, K3.1 Pro results use max reasoning effort, temperature 1.0 and top-p 0.95.
These are useful claims for selecting trial tasks, not independent proof of production superiority. Some comparison scores come from other vendors’ posts, some benchmarks are internal, and harnesses, fallback behavior and compute budgets differ. AccessAllGPT did not reproduce any score. Buyers should evaluate accepted outcomes, latency, token use and failure recovery on one fixed workload.
Open weight does not mean unrestricted
The custom Kimi K3.1-Pro License grants rights to use, copy, modify, deploy, fine-tune, distribute and sell the software and derivatives, subject to conditions. A licensee operating a model-as-a-service business whose aggregate affiliate revenue exceeds $20 million over any consecutive 12 months must reach a separate agreement with Moonshot before commercial use. A commercial product above 100 million monthly active users or $20 million monthly revenue must prominently display “Kimi K3.1-Pro.”
The license exempts defined internal use and use through Moonshot’s official products or certified inference partners from those two conditions. That is a material procurement boundary, not a legal conclusion. Teams should preserve the exact license version, map their service and affiliate revenue to its definitions, and obtain qualified advice before distributing or monetizing a derivative.
Choose managed access or weights for the right reason
Use the managed API for a fast capability trial, especially when the decision depends on output quality rather than infrastructure control. Measure both API tiers on identical prompts because Highspeed’s doubled token rates only make sense if its latency changes user or system outcomes. Record cache hits and writes separately; a nominal input rate is not an end-to-end task cost.
Consider the weights when offline operation, data control, model modification or infrastructure ownership justifies the engineering and license review. Do not choose self-hosting only because the repository exists. The 2.8T total footprint, custom code and sparse distributed serving path require capacity planning and a reproducible deployment before cost or sovereignty claims are credible.
What builders and buyers should do next
Trial kimi-k3.1-pro on a frozen set of coding, multimodal, long-context and tool-use tasks. Capture reasoning effort, prompt length, image handling, cache state, output tokens, latency, retries, factual errors and human acceptance. Add Highspeed only where response time is decision-relevant, and keep web search out of production while Moonshot’s own guide advises against it.
For weights, first validate the license, artifact hashes, supported runtime and a realistic memory and interconnect budget. Constrain deployment until representative quality, throughput and safety gates pass. Wait if the team lacks distributed-model operations or legal clarity. Reject the migration if the benchmark advantage disappears on the actual workload or total cost exceeds a managed alternative.
Copy-ready Kimi K3.1 Pro adoption record
Use one record per workload so API and self-hosted claims stay tied to a reproducible configuration.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Users, tasks, input modalities, context distribution, tools, volume, latency target and failure consequence.
kimi-k3.1-pro, kimi-k3.1-pro-highspeed or exact weight revision; region, provider, runtime and hardware.
Reasoning effort, streaming, images, structured output, tool definitions, max completion tokens, cache TTL and fixed settings.
Frozen prompts, expected behaviors, edge cases, long-context retrieval checks, tool failures and competitor baseline.
Acceptance, factual errors, tool success, latency percentiles, retries, input/output tokens, cache hits and cost per accepted task.
Artifact revision and hashes, license review, storage, accelerators, memory, interconnect, runtime, quantization and rollback.
Access, logging, data handling, evaluation cadence, human review, safety controls, incident owner and fallback.
Trial, constrain, wait or reject; chosen route, traffic ceiling, unresolved evidence, owner and retest date.
Primary sources
Browse the publication-wide evidence index →
- Kimi K3.1 Pro Tech Blog: Advancing Open IntelligenceMoonshot AI / Kimi · Reviewed: Structured metadata; launch overview; availability; coding, 3D and finance claims; cybersecurity release rationale; benchmark footnotes · Retrieved · Supports: Moonshot describes K3.1 Pro and Highspeed, their product and API availability, planned open-weight release, vendor benchmark results and stated open-weight safety rationale. Its structured data carries a 2026-10-24 dateModified value.
- Kimi K3.1 Pro API quickstartKimi API Platform · Reviewed: Overview; access requirements; model IDs; reasoning effort; streaming; vision; structured output; dynamic tools; context caching; official tools; important limits · Retrieved · Supports: The live API guide documents kimi-k3.1-pro and kimi-k3.1-pro-highspeed, a minimum $1 top-up, always-on thinking, 1M context, vision and tool behavior, fixed sampling settings and current web-search warning.
- Model Inference PricingKimi API Platform · Reviewed: K3 series model IDs, standard input and output rates, cached-input rates, cache-write rates, automatic cache billing and batch discount · Retrieved · Supports: Kimi lists Pro at $2 input and $10 output per million tokens and Highspeed at $4 input and $20 output, with separate cache prices. Standard Pro supports batch inference at 60% of standard price.
- moonshotai/Kimi-K3.1-Pro model card and repositoryMoonshot AI on Hugging Face · Reviewed: Repository creation record; model introduction; model summary; configuration; evaluation table; files and versions; license identifier · Retrieved · Supports: The public repository was created November 2, 2026 and contains 96 safetensor shards plus configuration and inference code. The model card identifies a 2.8T-parameter, 104B-active multimodal MoE with 1,048,576-token context and vendor-reported evaluations.
- Kimi K3.1-Pro LicenseMoonshot AI on Hugging Face · Reviewed: Granted rights; notice requirement; Model as a Service definition and revenue threshold; large-product attribution threshold; exemptions; warranty disclaimer · Retrieved · Supports: The custom license permits use, modification, deployment, fine-tuning and distribution, but introduces a separate-agreement condition for qualifying model-as-a-service businesses and an attribution condition for very large commercial products.
Limitations
This is a first-party documentation and artifact review, not hands-on model testing, an independent benchmark, a security assessment, legal advice or a serving-cost study. AccessAllGPT did not download or hash all 96 weight shards; instantiate the custom model code; verify compatibility with a production inference engine; measure storage, accelerator memory, interconnect, throughput, latency, energy or quality; call either API route; test the one-million-token window, vision, tool calling, structured output, caching or batch inference; reproduce Moonshot’s benchmark or safety results; inspect training data; compare K3.1 Pro against another model; verify authenticated product rollout, zero-data-retention terms or certified partners; or determine how the custom license applies to a particular business. Moonshot controls all cited surfaces, and capability, speed and benchmark statements are vendor claims.
Disclosures
AccessAllGPT received no weights, hosted access, API credits, private demo, briefing, benchmark data, legal review, payment or compensation for this article. Moonshot AI did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or certify Kimi K3.1 Pro. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Moonshot AI or Kimi. Publication-wide relationships are listed on the disclosures page.
