Terza Inference (SN103), the New Inference Player on the Block

Bittensor’s inference market includes several approaches to model delivery. Chutes (SN64) focuses on serverless compute for running models at scale, while Targon (SN4) emphasizes confidential, enclave-protected compute.

Engy (SN53) focuses on verifiable inference, while SayGM (SN28) aggregates model routes through an API gateway.

Together, these projects address different needs across compute availability, privacy, verification, and aggregation. Terza Inference enters this landscape with a marketplace approach, allowing independent miners to compete for requests through a shared gateway.

What Is Terza Inference?

Terza Inference (SN103) is an inference marketplace operating on Bittensor in early access. It connects developers to models served by independent miners through a single API key.

Its gateway selects an eligible supplier, checks reported usage against delivered output, and settles the actual charge.

Terza describes its proposition as “Many models in. One API out.” Developers access competing suppliers through a common interface, while miners set their own quotes and validators evaluate their service. The intended benefit is simpler access to model supply, with competition on price and checks on delivery.

How Requests Move Through Terza

Terza connects supplier selection, output delivery, usage verification, and payment within one workflow. Requests pass through five main stages.

1. A developer submits a request using a scoped API key, and the maximum possible cost is reserved from their prepaid USDC, USDT, or TAO balance before processing begins.

2. The cheapest healthy offer is selected from approved miners with valid quotes and on-chain registration, unless the developer chooses a specific offer.

3. The selected miner processes the request through its own upstream infrastructure, accepting only Ed25519-signed gateway requests and rejecting replayed requests or those older than 60 seconds.

4. After delivery, the miner’s reported usage is compared with the request and actual output, with implausible claims rejected and the reason recorded.

5. The verified cost is charged and the unused reservation released, while failed requests and rejected receipts incur no charge.

Independent validators assess available offers through paid probes designed to resemble customer traffic. Their evaluations inform the weights submitted on-chain, connecting miner rewards to measured service.

How Competing Miners Quote Work

Multiple miners can offer the same model at different prices, giving developers a choice of suppliers through one interface. Each listing on Terza’s pricing page represents a miner’s quote for a particular model and service tier, with information to help assess cost and reliability.

1. Pricing is displayed in US dollars per million tokens for text models, while media workloads use the relevant billing unit.

2. Percentage comparisons show how each quote differs from the model provider’s official list price.

3. The 24-hour success rate measures requests completed through Terza during that period.

4. Upstream uptime, where provided, reflects the miner’s own monitoring rather than Terza’s measurement of completed requests.

Miners set their own quotes, which can change and are checked again when requests are routed. The final charge uses the quote of the supplier that actually serves the request.

Developers can also adjust how suppliers are selected, depending on their priorities.

1. Service tiers offer economy, standard, and priority options.

2. Routing preferences prioritize the lowest price, highest success rate, fastest response, or a balance of these factors.

3. Offer pinning lets developers choose a specific supplier.

4. Automatic failover allows another offer to be tried when an attempt fails, without charging for the failed attempt.

Together, these controls let developers consider cost alongside service performance, although the available choices depend on the supply listed for each model.

How Validators Assess Miner Offers

Terza’s scoring framework requires miners to meet minimum quality standards before competing for rewards on price. Validators assess each offer against the following requirements.

1. Accuracy must reach at least 75%, supported by sufficient recent probe evidence.

2. Probe availability must reach at least 90%.

3. Rejected customer receipts must account for no more than 1% of customer receipts over a 24-hour period.

4. Model behaviour is checked against calibrated references to assess whether the supplied model matches the advertised identity.

5. Media outputs must meet additional availability and conformance requirements.

Offers that fail any required check score zero, while qualifying offers compete on price, with cheaper quotes receiving greater weight up to a threefold cap. Validators submit these weights on-chain to inform emissions distribution.

One Interface Across Different Workloads

Terza supports several AI workloads through its common interface. Developers can point compatible OpenAI or Anthropic clients at Terza, retaining familiar request structures.

The supported categories include the following.

1. Text and reasoning through chat completions, Responses, and Messages interfaces.

2. Retrieval through embeddings and reranking.

3. Images through image-generation endpoints.

4. Audio through speech, transcription, and realtime voice.

5. Asynchronous media through tasks such as video and music generation.

Usage checks depend on the workload. For text, the gateway compares reported consumption with the request and delivered text. For media, charges are based on measured characteristics such as duration, resolution, and count.

This structure is intended to reduce the integration work involved in accessing multiple suppliers. Developers maintain one gateway connection while choosing from the models and offers available behind it.

Participation During Early Access

Terza provides three routes into the network, depending on whether participants want to use models, supply them, or evaluate their delivery.

1. Builders create an API key, fund a prepaid balance, and submit requests to listed models.

2. Miners request an invitation, register a Bittensor hotkey, submit models and prices for review, and maintain a public HTTPS endpoint. They provide their own upstream capacity and retain their upstream credentials.

3. Validators read the signed public offer directory, test services through paid requests, and submit commit-reveal weights. They do not require access to Terza’s database or miners’ credentials.

Miner admission remains invitation-only during early access. The breadth of available supply will therefore depend partly on the operators admitted and the models they can reliably serve.

What Terza Still Has to Demonstrate

Terza brings competing miner offers, usage verification, and validator evaluation into one inference marketplace. Its longer-term appeal will depend on whether this structure delivers meaningful savings and reliable service as demand grows.

A competitive quote is useful only when the requested model is available and delivers suitable output. Receipt checks help verify billed usage, but developers must still assess response quality and accuracy for their applications.

For Bittensor, Terza offers a shared access point where independent suppliers compete under defined delivery and evaluation rules. The next test is whether that competition can support dependable service beyond early access.

Useful Resources

Follow Terza Inference (SN103) Via:

Website: https://terza.si/

Docs: https://terza.si/docs/

X: https://x.com/terzaiference

Discord: https://discord.com/channels/799672011265015819/1560711080404189267