Phase 1 · PUBLIC SANDBOX LIVE Forked-Akash PoS chain with an EVM and a Provider Agent for full hardware utilization: Console deploys plus idle GPU/CPU mining into ABA. The public sandbox is live end-to-end: Console deploys, wallet, explorer, ABA/USDC DEX with real Noble USDC over IBC, one-click USDC deposits from Polygon & other chains, a desktop app with one-click mining, and on-chain buyback payouts. Mainnet follows after an audit and external validator onboarding. View delivery statusStatus
Product · Planned
Phase 3 PLANNED

One base URL. Visible compute.

The same idea as a managed inference API like AkashML, on our own chain: keep familiar OpenAI request shapes and open-weight models, but expose the job and ABA settlement behind each completion. The API is not a separate product from the Console; it is the programmatic way into it.

Designed for existing tools.

from openai import OpenAI

client = OpenAI(
  base_url="https://api.abakos.ai/v1",
  api_key="abk_..."
)

response = client.embeddings.create(
  model="qwen-embed",
  input=batch
)

Preview only. API keys and models are not issued yet. Billing is planned in ABA.

01

Request

Use an OpenAI-compatible endpoint.

02

Job

The gateway creates a funded Console job in ABA.

03

Auto-match

The cheapest active, qualified provider is assigned automatically.

04

Compute

The assigned provider processes and proves the work.

05

Trace

The response includes a job reference visible in the explorer.

Familiar shapes, open weights.

Like AkashML, the API serves open-weight models (Llama, Qwen, DeepSeek and similar) behind OpenAI-compatible endpoints. Closed frontier models are never falsely claimed. Pricing is metered per request and settled in ABA rather than a fiat balance.

Phase 3 PLANNED

Embeddings

The first API workload after Console settlement works.

Phase 3 PLANNED

Batch completions

Open-model completions with asynchronous job tracking.

Phase 4 PLANNED

Streaming chat

Ships together with Abakos Chat, not earlier: both need the same real-time job path.

Multi-turn requests, across different providers

Nothing changes on your side: send the full messages array each call, exactly like any other OpenAI-compatible provider. Underneath, each call can be matched to a different GPU provider without losing context, because the job commits to the exact context used instead of relying on any one provider to remember it. Full design in the technical docs (spec-session-continuity.md). Planned, not live yet.

Join the Abakos list.

Updates on all progress and products - sandbox, Console, API and Chat. For the fastest small updates and quick answers, hop into our Discord. Never token-price hype.

You'll get a short confirmation email. One list, unsubscribe any time. See the privacy notice.