Moonshot AI launched Kimi K3.1 Pro on October 23 through its apps and API, and promised the weights by November 4 (Kimi blog). It delivered on schedule. The public, ungated Hugging Face repository holds about 1.56 TB of weights across 96 shards (ArtificialWatch; Hugging Face).
Within days the same model was live on AkashML, Akash Network’s inference service (@akashnet), on Bittensor-linked routing via SayGM (SN28) (@say_gm_), and inside a trusted execution environment via KubeTEE (SN90) (SayGM TEE listing). Centralized players moved just as fast: NVIDIA lists it on build.nvidia.com and Cerebras demoed it (NVIDIA; @cerebras).
For AI-crypto, this is a useful real-world test. When a near-frontier model is free to download, the model stops being the moat. Hosting it becomes the product, and decentralized compute networks have to win on price, privacy, reliability, and distribution. Here’s what we can verify so far.
The model, briefly
From Moonshot’s model card and launch blog:
- Architecture: mixture-of-experts, 2.8T total parameters, 104B active per token, 896 experts with 16 selected per token.
- Context: 1,048,576 tokens. Inputs are text and images.
- Precision: MXFP4 weights with MXFP8 activations, trained that way (quantization-aware), so the released weights are what Moonshot itself serves.
- Recommended engines: vLLM, SGLang, or TokenSpeed.
- Positioning: post-trained from Kimi K3 for coding, 3D creation, financial research, and agentic work. On Terminal-Bench 4.0, Moonshot reports 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. These are Moonshot’s own reported numbers, run with its stated harnesses.
- Safety choice: Moonshot says it removed cyber-related training data for the open-weight release. That cut pass@3 on its internal exploit benchmark from 90.5% (without the measures) to 81.0%.
Two practical points matter more than the benchmarks for anyone hosting it.
It’s big. At 1.56 TB of weights, this is not a single-GPU model. Serving it means multi-GPU nodes with plenty of memory, which is exactly the kind of capacity decentralized GPU networks say they can aggregate.
The license has a revenue threshold. Per ArtificialWatch’s comparison, the Kimi K3.1 Pro license matches Kimi K3’s. Use, modification, and redistribution are free, but a company selling model access as a service with more than $20 million in revenue over any 12 months needs a separate agreement with Moonshot (ArtificialWatch; GIGAZINE). That’s far above what any decentralized host we cover reports today, but it is a ceiling to keep in mind.
What it costs, and where
Moonshot’s list price for K3.1 Pro is $2.00 per million input tokens and $10.00 per million output tokens (SayGM price comparison; AI Lounge). Here is what we could confirm on pages we opened this morning (Nov 7):
| Where | Input / 1M | Output / 1M | Notes |
|---|---|---|---|
| Moonshot (list) | $2.00 | $10.00 | Reference price |
| AkashML (Akash) | $1.75 | $10.00 | Cache read $0.18; labeled “Limited Trial” (AkashML) |
| SayGM, best listed rate | $1.29 | $6.44 | 35.6% below list; 19 providers listed (SayGM) |
| SayGM, recent effective (blended) | $1.43 | $7.26 | What buyers actually paid across providers |
| SayGM TEE route (KubeTEE, SN90) | $1.75 | $8.76 | Prompts sealed from host and operators, per the team (SayGM TEE) |
| NVIDIA build.nvidia.com | Free endpoint | — | Page says preview/playground only, not programmatic access (NVIDIA) |
Three takeaways:
- Decentralized routes launched below list. SayGM went live on Oct 30 at 27.9% below Moonshot’s price (@say_gm_). By this morning its best listed rate was 35.6% below. When many independent operators serve the same open weights, price competition starts on day one.
- AkashML is priced close to list. Its input price is 12.5% below Moonshot’s and output is the same. AkashML is betting on a different pitch: a drop-in endpoint compatible with the OpenAI and Anthropic SDKs, built on Akash’s GPU supply (@akashnet).
- Privacy can carry a price. The TEE route costs more than SayGM’s cheapest route but less than list. That suggests some buyers will pay for verifiable confidentiality, a theme Render’s proposed Forth gateway is also chasing (RNP-024).
The table doesn’t tell you cost per accepted result. As we argued in our piece on inference costs, a cheaper token that needs more retries, fails more often, or adds latency to an agent loop can cost more in practice. SayGM publishes some of this: 89.4 tokens/s throughput, 1.5 s time to first token, and a 100% request success rate over its recent measurement windows (SayGM). We didn’t find comparable public numbers for AkashML.
Why this matters for Akash
Akash’s pitch has always been that an open marketplace can supply GPUs more cheaply than hyperscalers. AkashML is the layer that turns that supply into something a developer can call with an API key, without renting and managing GPUs.
That matters because Akash’s on-chain demand has been flat-to-down lately. DefiLlama’s Akash fee series, which tracks lease fees paid by users, shows about $288K in both August and September 2026, and about $207K in October (DefiLlama API, retrieved Nov 7). Hosting the most talked-about open model of the week is a straightforward way to draw developers in. The open question is whether AkashML’s usage shows up as lease demand on the marketplace, which is what matters for providers and AKT, or whether it stays a separately reported service line.
The timing overlaps with Akash’s planned move to Solana. Prop 346 authorized an SPL AKT and a one-way migration portal, but no marketplace cutover date (our explainer). Real API customers on AkashML are a better test of Akash’s product than any chain migration. For provider-side economics, see our Akash provider breakdown.
Why this matters for Bittensor
On Bittensor, the K3.1 Pro rollout shows the incentive design doing what it’s meant to do. Subnets that serve or route inference added the model quickly, and miners competed on price:
- SayGM (SN28) listed it Oct 30 at 27.9% below list, routed through a verifiable TEE (@say_gm_).
- KubeTEE (SN90) serves it inside a TEE through SayGM (Ventura Labs).
- Terza Inference (SN103) said it tested Cursor’s Agent-mode tool-calling loop on kimi-k3.1-pro, among other models (Ventura Labs).
This links to the question we raised in our Bittensor deep dive: who pays for subnets? Gamma, Bittensor’s on-chain system of dollar-valued credits minted by burning subnet alpha, is now in the mainnet runtime (spec 488). It still starts disabled and needs both network-wide and per-subnet enablement (Subtensor Release 487 notes). If inference subnets switch it on, popular open models like K3.1 Pro are the kind of paid usage that could show up on-chain instead of in off-chain billing.
The risk for every decentralized host
Open weights cut both ways. The same release that lets Akash and Bittensor operators serve a near-frontier model also lets every centralized cloud do it, often with a free tier. NVIDIA’s page lists a free endpoint and 15K API calls in the last 30 days (NVIDIA). Cerebras is showcasing its speed, with an online-shopping agent demo that finished in 11.5 seconds (@cerebras).
When the model is a commodity, the durable advantages are:
- Price at equal reliability. That’s cost per accepted result, not list price.
- Privacy guarantees users can verify, such as TEEs and identity-separated routing.
- Distribution. Being the default endpoint inside coding tools, agent frameworks, and routers.
Decentralized networks have a real shot at the first two. The third is where they’ve historically been weakest.
What to watch
- AkashML usage disclosures. Any token volumes or paying-customer counts for K3.1 Pro, and whether they appear in Akash lease fees.
- AkashML pricing. It’s labeled “Limited Trial” at $1.75/$10.00. Does pricing move toward the sub-list rates on Bittensor routes once the trial ends?
- Gamma enablement. Whether root and any inference subnet turn on Gamma, which would put paid model usage on-chain.
- Reliability data. Published latency, throughput, and success rates from decentralized hosts, not just price.
- Kimi K3.1 Pro Highspeed. Moonshot hasn’t said whether the 2–3x faster variant is a separate checkpoint, and no Highspeed repository is public (ArtificialWatch). If it ships as open weights, hosts will need to serve it too.
Research and commentary only, not financial advice. Prices and figures are from the linked pages as of Nov 7, 2026 and change often.