Muna
Try Muna on us. Starter credit covers your first requests, then one key reaches all 3 Muna models, with the privacy posture shown on every route.
Munamuna
Open-weight chat, vision and reasoning models served on Muna's own inference stack.
These privacy labels describe Muna, the upstream model provider. Confidential requires all three: verified provider compute, end-to-end encryption and explicit zero data retention. Account and billing metadata are outside the content-retention claim.
| Provider | Muna |
|---|---|
| API operator country | US Legal home of the API operator, not an inference-location guarantee. |
| Headquarters | Not verified |
| Routing status | Active |
| Provider website | https://muna.ai/ |
| Models | 3 public models |
| Credits routes | 3 |
| Zero data retention | no |
| Verified confidential inference | Not verified |
| Policy note | Muna (NatML Inc.) operates its own serving stack on rented US GPUs, without reselling downstream model APIs. No training on customer content. Its published zero-retention policy permits incidental content fragments in diagnostic logs for up to 30 days, so these routes remain Standard, not strict ZDR or confidential/E2EE. Shared prefix and decoded-image caches are memory-only; prefix caching is not account-isolated. Request and edge metadata remain for 30 days, billing records for seven years. Renewable electricity coverage is unknown. Tools support auto/none selection; required tool selection is not currently supported. Policy source |
Documentation and policy sources
Inference locations
Provider API declarations are also available on each model's endpoint API. Direct chat and Responses replies separate advertised availability from serving location in usage.inference_location; the actual serving region remains unknown when not reported.
usage.provider_usage.gateway_region identifies the TrustedRouter gateway, not the provider's inference location. A value such as europe-west4 does not mean the model ran in Europe. usage.inference_location.serving_region remains null when the provider has not reported it. Older gateways may return the gateway-only field as provider_usage.region.
Where the upstream GPUs execute the model, not the provider's headquarters, API ingress, or data storage. ZDR and confidential inference do not by themselves restrict geography.
- Declared countries or regions
- United States
Provider-declared US inference on rented dedicated GPUs operated by Muna; no downstream model APIs. - Can locations change?
- Dynamic per-request placement across US nodes; provider declares no failover outside the US.
- Provider region pinning
- No per-request region pin or actual serving-region response field.
- Pinning through TrustedRouter
- Select provider.only=[muna] with allow_fallbacks=false; no finer regional pin is available.
- Infrastructure declaration
- United States (provider-declared), dynamically placed; specific datacenter region is not exposed.
- Evidence
- Marketplace application reviewed 2026-10-15 and authenticated catalog declaring us; not attested location.. Reviewed .
Provider selection alone is not a region pin. TrustedRouter's US provider filter selects company jurisdiction, not GPU location. A regional gateway hostname does not pin a downstream GPU. For a binding residency requirement, confirm the model, region, failover scope and enforced controls before sending sensitive data.
Measured performance
25 samplesContinuously sampled across Muna's routed models: p50 TTFT, effective throughput, and success rate. Effective throughput uses provider-reported output tokens over complete request time. Unsupported route and probe-configuration rows are separated from provider downtime. No prompt or output content stored.
| p50 TTFT | 1288 ms |
|---|---|
| Effective throughput | 138 tok/s n=1 |
| Uptime | 100.00% |
| Model | p50 TTFT | Effective throughput | Uptime | Config excluded | Availability samples |
|---|---|---|---|---|---|
| qwen/qwen3.8-27b | 1284 ms | 138 tok/s n=1 | 100.00% | 1 router_error |
11 |
| orcarouter/qwen-3.8-27b-uncensored | 1832 ms | — | 100.00% | — | 10 |
| google/gemma-4-26b-a4b-it | 1288 ms | — | 100.00% | — | 4 |
Muna performance history · Full provider & model leaderboard.
Models served by Muna.
Each row links to pricing, provider, benchmark, and API pages for the model.
| Model | AI IQ | Context | Input | Cached input | Output |
|---|---|---|---|---|---|
google/gemma-4-26b-a4b-itGoogle: Gemma 4 26B A4B |
IQ 99#131 | 262,144 | $0.068575/1M | $0.0211/1M | $0.3165/1M |
orcarouter/qwen-3.8-27b-uncensoredqwen-3.8-27b-uncensored |
— | 262,144 | $0.36925/1M | $0.0844/1M | $2.90125/1M |
qwen/qwen3.8-27bQwen: Qwen3.8 27B |
IQ 115#57 | 65,536 | $0.26375/1M | $0.0211/1M | $1.84625/1M |
Questions
Does Muna have zero data retention?
TrustedRouter does not currently mark Muna as provider-level zero data retention. Use trustedrouter/zdr or provider.min_privacy=zdr to select a different eligible route, and review the linked policy source for changes.
Is Muna end-to-end encrypted?
TrustedRouter does not currently mark Muna as end-to-end encrypted at the provider boundary. The TrustedRouter gateway is still attested, but the selected provider normally receives the request in order to run the model. Use trustedrouter/e2e for the stronger route requirement.
Which Muna models are available through TrustedRouter?
This page currently lists 3 public Muna models, with live pricing, route count, context length, measured performance when available, and links to each model's provider and benchmark pages.