# Partition Labs — full reference

> Flat-rate LLM inference. A subscription buys a guaranteed share of GPU capacity on one
> pool, plus prepaid pay-as-you-go credits, behind an OpenAI- and Anthropic-compatible API.

A full share is the GPU load of one agent running continuously at concurrency 1, sold in 24 share units. Reserved capacity is guaranteed under contention; nobody is rejected for volume. Service tiers set queue admission order, never stream speed. Prefix-cache writes are free.

## Base URLs

```
OPENAI_BASE_URL=https://api.partitionlabs.ai/v1
ANTHROPIC_BASE_URL=https://api.partitionlabs.ai
```

Streaming, tool calls and image input work on both. Files and batches use the OpenAI
shapes (`/v1/files`, `/v1/batches`). Pass `service_tier` on a request (OpenAI: top-level;
Anthropic: the same field); the tier that served it is echoed back in `usage`.

## Subscriptions

Prices are per month in USD. A subscription reserves share on ONE pool; every pool runs one
model, resident in memory. Slots, queues and waits are per API key and are what predict a 429.

### Qwen 3.8 27B FP8 — 4k context pool (`qwen3.8-27b-fp8-4k`)

4k context. One full share is one always-on 4k request; a share unit is priced at $10/month because F(4k) is what the card actually holds at that length.

- Solo: $10.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 4k context; batch 2 concurrent (24h) / 4 (7d)
- Dev: $40.00/month; 4 of 24 share units; 4 active + 4 flex concurrent; queues 8 standard / 16 flex; max wait 60s standard / 240s flex, then 429; up to 4k context; batch 8 concurrent (24h) / 16 (7d)
- Half share: $120.00/month; 12 of 24 share units; 12 active + 12 flex concurrent; queues 24 standard / 48 flex; max wait 60s standard / 240s flex, then 429; up to 4k context; batch 24 concurrent (24h) / 48 (7d)
- Full share: $240.00/month; 24 of 24 share units; 24 active + 24 flex concurrent; queues 48 standard / 96 flex; max wait 60s standard / 240s flex, then 429; up to 4k context; batch 48 concurrent (24h) / 96 (7d)

### Qwen 3.8 27B FP8 — 8k context pool (`qwen3.8-27b-fp8-8k`)

8k context. One full share is one always-on 8k request; a share unit is priced at $12/month because F(8k) is what the card actually holds at that length.

- Solo: $12.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 8k context; batch 2 concurrent (24h) / 4 (7d)
- Dev: $48.00/month; 4 of 24 share units; 4 active + 4 flex concurrent; queues 8 standard / 16 flex; max wait 60s standard / 240s flex, then 429; up to 8k context; batch 8 concurrent (24h) / 16 (7d)
- Half share: $144.00/month; 12 of 24 share units; 12 active + 12 flex concurrent; queues 24 standard / 48 flex; max wait 60s standard / 240s flex, then 429; up to 8k context; batch 24 concurrent (24h) / 48 (7d)
- Full share: $288.00/month; 24 of 24 share units; 24 active + 24 flex concurrent; queues 48 standard / 96 flex; max wait 60s standard / 240s flex, then 429; up to 8k context; batch 48 concurrent (24h) / 96 (7d)

### Qwen 3.8 27B FP8 — 16k context pool (`qwen3.8-27b-fp8-16k`)

16k context. One full share is one always-on 16k request; a share unit is priced at $14/month because F(16k) is what the card actually holds at that length.

- Solo: $14.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 16k context; batch 2 concurrent (24h) / 4 (7d)
- Dev: $56.00/month; 4 of 24 share units; 4 active + 4 flex concurrent; queues 8 standard / 16 flex; max wait 60s standard / 240s flex, then 429; up to 16k context; batch 8 concurrent (24h) / 16 (7d)
- Half share: $168.00/month; 12 of 24 share units; 12 active + 12 flex concurrent; queues 24 standard / 48 flex; max wait 60s standard / 240s flex, then 429; up to 16k context; batch 24 concurrent (24h) / 48 (7d)
- Full share: $336.00/month; 24 of 24 share units; 24 active + 24 flex concurrent; queues 48 standard / 96 flex; max wait 60s standard / 240s flex, then 429; up to 16k context; batch 48 concurrent (24h) / 96 (7d)

### Qwen 3.8 27B FP8 — General pool (`qwen3.8-27b-fp8`)

32k context. One full share is one always-on 32k request; a share unit is priced at $20/month because F(32k) is what the card actually holds at that length.

- Solo: $20.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 32k context; batch 2 concurrent (24h) / 4 (7d)
- Duo: $40.00/month; 2 of 24 share units; 2 active + 2 flex concurrent; queues 4 standard / 8 flex; max wait 60s standard / 240s flex, then 429; up to 32k context; batch 4 concurrent (24h) / 8 (7d)
- Dev: $80.00/month; 4 of 24 share units; 4 active + 4 flex concurrent; queues 8 standard / 16 flex; max wait 60s standard / 240s flex, then 429; up to 32k context; batch 8 concurrent (24h) / 16 (7d)
- Pro: $160.00/month; 8 of 24 share units; 8 active + 8 flex concurrent; queues 16 standard / 32 flex; max wait 60s standard / 240s flex, then 429; up to 32k context; batch 16 concurrent (24h) / 32 (7d)
- Heavy: $240.00/month; 12 of 24 share units; 12 active + 12 flex concurrent; queues 24 standard / 48 flex; max wait 60s standard / 240s flex, then 429; up to 32k context; batch 24 concurrent (24h) / 48 (7d)
- Max: $480.00/month; 24 of 24 share units; 22 active + 24 flex concurrent; queues 44 standard / 96 flex; max wait 60s standard / 240s flex, then 429; up to 32k context; batch 48 concurrent (24h) / 96 (7d)

### Qwen 3.8 27B FP8 — 64k context pool (`qwen3.8-27b-fp8-64k`)

64k context. One full share is one always-on 64k request; a share unit is priced at $30/month because F(64k) is what the card actually holds at that length.

- Solo: $30.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 64k context; batch 2 concurrent (24h) / 4 (7d)
- Dev: $120.00/month; 4 of 24 share units; 2 active + 2 flex concurrent; queues 4 standard / 8 flex; max wait 60s standard / 240s flex, then 429; up to 64k context; batch 4 concurrent (24h) / 8 (7d)
- Half share: $360.00/month; 12 of 24 share units; 6 active + 6 flex concurrent; queues 12 standard / 24 flex; max wait 60s standard / 240s flex, then 429; up to 64k context; batch 12 concurrent (24h) / 24 (7d)
- Full share: $720.00/month; 24 of 24 share units; 12 active + 12 flex concurrent; queues 24 standard / 48 flex; max wait 60s standard / 240s flex, then 429; up to 64k context; batch 24 concurrent (24h) / 48 (7d)

### Qwen 3.8 27B FP8 — 128k context pool (`qwen3.8-27b-fp8-128k`)

128k context. One full share is one always-on 128k request; a share unit is priced at $50/month because F(128k) is what the card actually holds at that length.

- Solo: $50.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 128k context; batch 2 concurrent (24h) / 4 (7d)
- Dev: $200.00/month; 4 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 128k context; batch 2 concurrent (24h) / 4 (7d)
- Half share: $600.00/month; 12 of 24 share units; 3 active + 3 flex concurrent; queues 6 standard / 12 flex; max wait 60s standard / 240s flex, then 429; up to 128k context; batch 6 concurrent (24h) / 12 (7d)
- Full share: $1200.00/month; 24 of 24 share units; 6 active + 6 flex concurrent; queues 12 standard / 24 flex; max wait 60s standard / 240s flex, then 429; up to 128k context; batch 12 concurrent (24h) / 24 (7d)

### Qwen 3.8 27B FP8 — 256k context pool (`qwen3.8-27b-fp8-256k`)

256k context. One full share is one always-on 256k request; a share unit is priced at $90/month because F(256k) is what the card actually holds at that length.

- Solo: $90.00/month; 1 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 256k context; batch 2 concurrent (24h) / 4 (7d)
- Dev: $360.00/month; 4 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 256k context; batch 2 concurrent (24h) / 4 (7d)
- Half share: $1080.00/month; 12 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 256k context; batch 1 concurrent (24h) / 2 (7d)
- Full share: $2160.00/month; 24 of 24 share units; 1 active + 1 flex concurrent; queues 2 standard / 4 flex; max wait 60s standard / 240s flex, then 429; up to 256k context; batch 1 concurrent (24h) / 2 (7d)

Hours are a floor, counted against a fully sold pool. Reserved capacity is guaranteed under contention; additional usage is available on a best-effort basis, so a quiet pool potentially gives you more than your share. Interactive work is bursty rather than continuous — an hour of non-stop processing typically covers two to four hours of interactive hands-on development.

Max is the GPU load of one always-on agent. If you need several, or a pool nothing else runs on, we size it with you rather than from a table. Contact: sales@partitionlabs.ai.

## Pay as you go

Prepaid credits, priced per million tokens. Same model and the same tokens per second as a
subscription — what differs is admission.

- Priority (`priority`, queue-jump under contention): input $0.80, cache write free, cached read $0.10, output $6.00
- Standard (`standard`, the default): input $0.40, cache write free, cached read $0.05, output $3.00
- Flex (`flex`, same model and speed; last in line): input $0.30, cache write free, cached read $0.0375, output $2.25
- Batch (`batch`, async; completes within 24h or 7d): input $0.20, cache write free, cached read $0.025, output $1.50

Pay-as-you-go is best-effort: it fills capacity that reserved tenants are not using, which is most of the time. Under contention a subscriber stays ahead of it — but only inside the share they bought. Past that, their traffic falls below every paying pay-as-you-go request, because the reservation is spent. Nobody is rejected for volume; whoever ranks lower queues.

No context surcharge at launch.

Batch finishes inside the window you pick — 24 hours or 7 days, same price. We refuse work we cannot schedule in time rather than accept it and run late; anything still unfinished at the deadline expires and is not billed. Results stay available for 30 days after completion, or less if you ask. Files up to 512 MiB, batches up to 100,000 requests or 256 MiB; storage is included with every plan.

Credits are bought in packages starting at $5.00. Credits expire 12 months after purchase.

Pay-as-you-go requests run at up to 64k context, with a maximum queue wait of 60s (standard) or 240s (flex), then 429.

## Limits

- Max output tokens per request: 32,768
- Max request body: 32 MiB
- Images per request: 20, up to 33,177,600 pixels in total
- Max request wall clock: 1800s; a stream that stalls for 120s is aborted; keepalives every 15s
- Files up to 512 MiB; batches up to 100,000 requests or 256 MiB

## Questions

### What is a guaranteed GPU share?

A full share is the GPU load of one agent running continuously at concurrency 1 on one pool. It is sold in 24 share units, and each tier holds a fixed number of them: the Duo tier holds 2 of 24, which is at least 2 hours a day of non-stop processing, counted against a fully sold pool. We never sell more share units than a pool can serve, and the ledger that proves it is on the homepage.

### What happens when I use more than my share?

Nothing is rejected for volume. Under contention a subscriber inside their share goes ahead of every pay-as-you-go request; past it, their traffic ranks below paying pay-as-you-go requests and queues until they are back under share. A 429 comes only from the per-key limits: every tier publishes its concurrent slots, its queue depth and a maximum wait, and a request past all three is refused rather than kept waiting.

### Do service tiers change speed?

No. Every tier runs the same model at the same tokens per second; a tier sets queue admission order, so a higher tier starts sooner when the pool is busy. The tiers are Priority (queue-jump under contention), Standard (the default), Flex (same model and speed; last in line), Batch (async; completes within 24h or 7d). Pass service_tier on a request (OpenAI: top-level; Anthropic: the same field) and the tier that served it is echoed back in usage.

### Does it work with the OpenAI and Anthropic SDKs?

Yes. Set OPENAI_BASE_URL to https://api.partitionlabs.ai/v1 or ANTHROPIC_BASE_URL to https://api.partitionlabs.ai and keep the client, the types and the retries you already have. Streaming, tool calls and image input work on both. Files and batches use the OpenAI shapes (/v1/files, /v1/batches).

### What do prefix-cache writes cost?

Nothing. Cache writes are free on every service tier, a byproduct of RadixAttention rather than a promotion. Cached reads are billed at the rate card's cached-read price, $0.05 per million tokens on Standard.

### What is the difference between a subscription and pay-as-you-go credits?

A subscription reserves share on one pool for a flat monthly price. Pay-as-you-go credits are prepaid, priced per million tokens, and best-effort: they fill the capacity reserved tenants are not using, which is most of the time. Packages start at $5.00, credits expire 12 months after purchase, and pay-as-you-go requests run at up to 64k context.

### What do "Sold out" and "Waitlist only" mean?

Sold out means every share unit on that pool is committed; we never oversell. Waitlist only means the pool is not taking new subscriptions right now, whatever its room. Either way, join the waitlist for the tier you want and we notify in the order people joined as capacity opens.

### Which models and context lengths are available?

Qwen 3.8 27B FP8 on the 4k context pool (model id qwen3.8-27b-fp8-4k, up to 4k context); Qwen 3.8 27B FP8 on the 8k context pool (model id qwen3.8-27b-fp8-8k, up to 8k context); Qwen 3.8 27B FP8 on the 16k context pool (model id qwen3.8-27b-fp8-16k, up to 16k context); Qwen 3.8 27B FP8 on the General pool (model id qwen3.8-27b-fp8, up to 32k context); Qwen 3.8 27B FP8 on the 64k context pool (model id qwen3.8-27b-fp8-64k, up to 64k context); Qwen 3.8 27B FP8 on the 128k context pool (model id qwen3.8-27b-fp8-128k, up to 128k context); Qwen 3.8 27B FP8 on the 256k context pool (model id qwen3.8-27b-fp8-256k, up to 256k context). Every pool runs one model, resident in memory, and a subscription is a share of one pool.

## Pages

- https://partitionlabs.ai/ — what a guaranteed GPU share is, and the live capacity ledger
- https://partitionlabs.ai/pricing — subscription ladder per pool, the pay-as-you-go rate card, and answers to common questions
- https://partitionlabs.ai/docs — quickstart: base URLs for the OpenAI and Anthropic SDKs, model ids, service_tier, files and batches, limits
- https://partitionlabs.ai/status — measured time-to-first-token per band, last 24 hours
- https://partitionlabs.ai/llms.txt — the short form of this file