The models everyone's asking for are already on 0G

Running a frontier model today means picking a provider, accepting its endpoint, and trusting its data policy on faith. Add a second model and you add a second provider, a second endpoint, and a second policy to accept blind. For a person in a chat window that gap is fine. For an autonomous agent working on proprietary code or user data, trust-by-policy is the weak link, and it multiplies with every model you bolt on.

The 0G Private Computer collapses that. One OpenAI-compatible endpoint, one key, and every model labeled by trust tier so you know what you can actually verify before a prompt leaves your machine. Two new sets are already live: Kimi K3, Moonshot's flagship, on Verified Routing, and the GPT-5.6 series, Sol, Terra, and Luna, on Standard Routing. And because several models arrived on 0G with only a tweet and no page of their own, this is also the running record of every model on pc.0g.ai, sorted by trust tier.

What is launching

Both new sets are reachable now at pc.0g.ai, on the same app-sk-* key flow that already serves the rest of the lineup. The API surface is the 0G Router at https://router-api.0g.ai/v1, a single public endpoint in the OpenAI format that normalizes requests and responses across every model on the platform.

Kimi K3: Moonshot's flagship, verified routing

Kimi K3 is Moonshot AI's newest flagship, a roughly 2.8-trillion-parameter open-weight Mixture of Experts model (16 of 896 experts active per token) with a native 1M-token context window. On 0G it runs through Verified Routing, TeeTLS-attested, straight to Moonshot's official API. This is a direct connection, not a reseller, and there is no markup: you pay Moonshot's official list price and 0G adds a verifiable routing layer on top. It takes multimodal input (text, image, and video) and returns text, keeps deep thinking on at all times, and supports function calling and implicit prompt caching.

GPT-5.6: Sol, Terra, and Luna, standard routing

The GPT-5.6 series comes in three sizes: Sol, the flagship for advanced reasoning and complex coding; Terra, the balanced tier; and Luna, the cost-efficient tier for high-volume workloads. All three are reasoning models with tool use, take text and image input, return text, and carry a 1M-token context window with up to 128K tokens of output.

The GPT-5.6 series runs on Standard Routing. That means full access to the models through 0G's OpenAI-compatible endpoint, with $0G settlement across the same key flow as everything else, but no TEE attestation. It is not a verifiability claim, and the deployment does not pretend otherwise. If your workload needs an attested path, the Verified and Private tiers below are where to look. If you want the model plus one endpoint for your whole stack, Standard Routing is the door.

Kimi K3 supports cached input at $0.30 per million tokens. OpenAI does not disclose parameter counts for the GPT-5.6 series; Sol, Terra, and Luna are capability tiers (flagship, balanced, cost-efficient). For the GPT-5.6 series, requests over 272K input tokens move to a long-context rate (input at two times the standard, output at one-and-a-half times); on Sol that is $9.00 input and $40.50 output per million.

At max reasoning effort (July 24, 2026), GPT-5.6 Sol scores 58.9 and Kimi K3 scores 57.1 on the Artificial Analysis Intelligence Index, among the highest of any model reachable on 0G.

The three trust tiers

The reason the 0G Private Computer labels every model is that "which model" and "how much you can verify" are two separate questions. 0G answers the second one with three trust modes, each backed by a different hardware guarantee (0G router privacy docs).

Private Inference (TeeML) seals the inference itself. The model runs inside a Trusted Execution Environment that 0G operates, Intel TDX with TEE-enabled GPUs; the prompt enters the enclave encrypted, never leaves it, and the operator cannot read it. This is the sovereign tier, the right one for workloads that legally need inference where no one, including 0G, can see the input.

Verified Routing (TeeTLS) attests the path rather than sealing the inference. 0G's broker, itself running inside a TEE, relays your request to the upstream provider over attested TLS, so the response provably came from the real model and the transport cannot be intercepted or silently redirected. The inference runs at the upstream under its own data policy. Kimi K3 sits here, with Moonshot's official API as the upstream, and so does Hunyuan 3, Tencent's flagship, routed through Tencent Cloud's official gateway.

Standard Routing gives you the model through 0G's OpenAI-compatible endpoint with $0G settlement, and no independent verifiability method. Full model access and single-endpoint convenience, but not a verifiability claim. The GPT-5.6 series sits here. Some frontier models are not reachable through an attestable path yet; Standard Routing is how 0G still offers them without overstating what they are.

Every model on 0G, by trust tier

Several models reached 0G with only an announcement post and never got a page of their own. This section fixes that. It is the running record of every model live on pc.0g.ai today, 23 in all, sorted by the trust tier each one runs on. Where a model already has a full write-up, it is linked.

Private Inference (TeeML), sealed enclave, the operator cannot read the prompt:

• 0GM-1.0-35B-A3B, 0G's own Apache 2.0 model, trained on 0G's decentralized GPU network, inference sealed inside a 0G-operated enclave. It outperforms its Qwen 3.6 base on 12 of 14 MMLU-Pro subjects and on AIME 2026 (model card). Announcement
• 0GM-1.0-35B-A3B-SIA, an alignment-tuned variant of 0GM-1.0 built with Nanyang Technological University, using inference-time value steering to raise output quality.
• GLM-5.2, Z.ai's open-weight flagship, added a sealed TeeML tier in July and is the default private model in 0G Private Chat. Announcement

Verified Routing (TeeTLS), the path to the upstream is hardware-attested and provable onchain:

• Kimi K3, the newest addition to the tier, Moonshot's flagship, routed direct to Moonshot's official API.
• Hunyuan 3, Tencent's 295B-parameter flagship (21B active) with a 256K context, routed through Tencent Cloud's official gateway at $0.13 per million input tokens and $0.53 per million output.
• DeepSeek V4 Pro, open-weights frontier reasoning, 1M context. Announcement
• GLM-5, Z.ai's 744B foundation and the first GLM on 0G. Announcement
• Qwen 3.7 Max, Alibaba's flagship with native function calling and web search, part of the wider Alibaba Cloud and Qwen family on 0G (Qwen3.7-Plus, Qwen3.6-Plus, and the Qwen3-VL vision model). Model and partnership
• MiniMax M3, a frontier natively-multimodal model, live on 0G since June and written down here for the first time.
• Also on this tier: GLM-5.1, Kimi K2.7-Code, and DeepSeek V4 Flash.

Standard Routing, the model through 0G's OpenAI-compatible endpoint, no attestation:

• GPT-5.6 Sol, Terra, and Luna, the newest additions to the tier.
• Claude Fable 5, Anthropic's 1M-context model, live on 0G since July and recorded here for the first time.
• Claude Opus 4.8 and Claude Sonnet 5, added the following day, also recorded here for the first time.

Beyond chat, 0G also serves Whisper Large v3 for speech-to-text and Z-Image for text-to-image, both inside sealed TeeML enclaves.

Build with it

Any OpenAI-compatible client runs against the 0G deployment with one base URL change. Swap the model id to target a different model on the same endpoint.

from openai import OpenAI

client = OpenAI(base_url="https://router-api.0g.ai/v1", api_key="app-sk-<YOUR_KEY>")
response = client.chat.completions.create(
    model="kimi-k3",  # or gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

Tool calling works end-to-end: send an OpenAI tools=[...] payload and the response returns a proper tool_calls array. For models on the Verified and Private tiers, set verify_tee: true in the request body to get onchain TEE attestation in the trace; the response returns a tee_verified field you can check against the provider's attestation onchain. Standard Routing models carry no attestation, so that field does not apply to them. Router Mode picks a model for the workload automatically; Advanced Mode lets you target a specific model id.

Pricing in $0G

Usage is USD-denominated and settled in $0G on 0G Chain, at an FX peg that floats hourly.

Kimi K3 is Moonshot's official list price with no 0G markup: $3.00 per million input tokens, $15.00 per million output, and $0.30 per million cached input.

The GPT-5.6 series is launch-priced 10% below the official rate. Per million input and output tokens: Sol at $4.50 / $27.00 (official $5.00 / $30.00), Terra at $2.25 / $13.50 (official $2.50 / $15.00), and Luna at $0.90 / $5.40 (official $1.00 / $6.00). Requests over 272K input tokens move to the long-context rate noted above.

Exact per-model prices, including cache rates and the long-context tiers, are on each model's page at pc.0g.ai.

Frequently asked questions

What models are live on 0G?

On Private Inference (TeeML, sealed inference): 0GM-1.0, 0GM-1.0-SIA, and GLM-5.2. On Verified Routing (TeeTLS, attested routing): Kimi K3, Hunyuan 3, DeepSeek V4 Pro, DeepSeek V4 Flash, GLM-5, GLM-5.1, Qwen 3.7 Max and the wider Qwen family (Qwen3.7-Plus, Qwen3.6-Plus, Qwen3-VL), MiniMax M3, and Kimi K2.7-Code. On Standard Routing (no attestation): the GPT-5.6 series (Sol, Terra, Luna) and the Claude family (Fable 5, Opus 4.8, Sonnet 5). 0G also serves Whisper for speech-to-text and Z-Image for text-to-image, both in sealed TeeML enclaves.

Is GPT-5.6 on 0G private or verifiable?

No. The GPT-5.6 series runs on Standard Routing, which gives you full access to the models through 0G's OpenAI-compatible endpoint with $0G settlement, but no TEE attestation. It is not a verifiability claim. For attested paths, use the Verified Routing or Private Inference tiers.

What is the difference between Private Inference, Verified Routing, and Standard Routing?

Private Inference (TeeML) runs the model inside a 0G-operated sealed enclave, so the operator cannot read the prompt. Verified Routing (TeeTLS) attests the routing path to an upstream provider, provable onchain, while the inference happens at the upstream. Standard Routing is the model through 0G's endpoint with no attestation.

How do I call Kimi K3 or GPT-5.6 on 0G?

Point any OpenAI-compatible client at https://router-api.0g.ai/v1, use an app-sk-* key from pc.0g.ai, and set the model id to kimi-k3, gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna.

What can Kimi K3 do?

Kimi K3 is Moonshot's flagship. It takes text, image, and video input, returns text, carries a 1M-token context window, keeps deep thinking on at all times, and supports function calling and implicit prompt caching. On 0G it runs through Verified Routing, direct to Moonshot's official API with no markup.

Where do I get a key?

Visit pc.0g.ai, connect your wallet, transfer $0G to the provider for the model you want, and generate an app-sk-* key. Plug it into your OpenAI SDK with base_url="https://router-api.0g.ai/v1"; the same Router endpoint serves every model on the Private Computer.

Build on 0G

• Try the new models now: pc.0g.ai
• Chat with models privately, no key: app.0g.ai/private-chat
• Read the trust-model details: docs.0g.ai
• Get $0G: get.0g.ai
• Follow @0G_labs for what goes live next on 0G private compute

Sources:

• 0G Router model list (live model ids, context windows, trust tiers, pricing)
• 0G router privacy and trust modes (Private / Verified / Standard definitions, TeeML and TeeTLS, verify_tee)
• Kimi K3 on the 0G Private Computer and GPT-5.6 Sol on the 0G Private Computer (specs, pricing, routing)
• Moonshot AI Kimi K3 (2.8T MoE architecture, 1M context) and GPT-5.6 pricing (official list rates)
• DeepSeek V4 Pro is live on 0G and GLM-5.2 is live on 0G (trust-tier companions)