Skip to content
Bastion Token Factory is now in preview — open-weight models through one OpenAI-compatible API. Join the waitlist →
Bastion Token Factory · Preview

Run open models at production scale

Access leading open models, optimize inference, choose how you deploy and adapt models with your own data — all through one platform, served on GPU capacity Bastion has secured in U.S. and EU data centers. Now in preview: join the waitlist for early access.

# pip install openai — change base_url and model
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.bastion-cloud.com/v1",
    api_key=os.environ["BASTION_API_KEY"],
)
resp = client.chat.completions.create(
    model="mistralai/Mistral-Small-3.2-24B-Instruct-2506",
    messages=[{"role": "user",
               "content": "Explain liquid cooling briefly."}],
)
print(resp.choices[0].message.content)

Discover leading open-weight models

Compare leading text, code and reasoning models, then choose the right fit for your workload. Explore all models

OOPENAI

gpt-oss-120b

An open-weight mixture-of-experts reasoning model with configurable reasoning effort and native tool use.

128K contextApache 2.0$0.15 / $0.60 per 1M
MSMICROSOFT

Phi-4

A compact 14B model with strong reasoning on math, science and code — efficient enough for high-volume workloads.

16K contextMIT$0.05 / $0.15 per 1M
MMISTRAL AI

Mistral Small 3.2

A fast, efficient European model with vision support and strong instruction following and function calling.

128K contextApache 2.0$0.08 / $0.24 per 1M

From model access to production in one platform

Everything you need to serve, scale and improve open models — without stitching together separate tools.

Optimized inference

Run open models on serving stacks tuned for latency, throughput and cost — with speculative decoding, prompt caching and cache-aware routing.

Scalable infrastructure

Move from early testing to sustained production traffic on B300 and GB300 capacity Bastion has locked in across U.S. and EU data centers.

Post-training

Adapt open models with your own data using LoRA, full fine-tuning or DPO, then deploy the resulting model on Token Factory.

Choose how you deploy

Start quickly with serverless inference, or choose dedicated capacity for greater production control.

Serverless inference

Explore models and start building through a familiar API, without planning infrastructure upfront.

Best for: evaluation, prototyping and variable workloads.

Preview pricing from $0.03 per 1M tokens · Batch −50%

Dedicated endpoints

Run models on isolated B300 or GB300 capacity when you need greater control over performance, scaling, model configuration or regional placement.

Best for: sustained traffic, custom models and production-critical workloads.

Per GPU-hour · reserve capacity now

Built for real production workloads

Put open models to work across high-value applications.

Coding agents

Build code generation, review, debugging and agentic software-development experiences.

AI search & RAG

Build agents that retrieve, verify and reason over your documents with embeddings and long context.

Fraud and risk

Investigate complex signals, support risk workflows and keep human review in the loop.

Customer support

Resolve tickets with fine-tuned assistants that follow your policies and tone.

Image & audio

Generate images with FLUX.1 [schnell] and transcribe speech with Whisper.

Evals & synthetic data

Run offline evaluation, labeling and data generation at half price with the Batch API.

Model catalog & preview pricing

Every model in the catalog is open-weight and released under a permissive license (Apache 2.0 or MIT), shown in the License column. All are served through an OpenAI-compatible API.

ModelProviderCategoryContextLicensePreview price (USD / 1M tokens)Highlights
Loading catalog…

Base: cost-efficient serving for most workloads. Fast: speculative decoding on reserved high-throughput capacity, at 2× the base price. Indicative preview prices, subject to change at general availability. Use of each model is subject to its license and acceptable-use policy. Additional open-weight models can be served on dedicated endpoints on request, subject to license review.

Estimate your monthly bill

Model
Input tokens / month—
Output tokens / month—
Serving flavor
Estimated monthly cost
$0
Input—
Output—
Rate (in / out per 1M)—
Mode—

Estimate at list prices, excluding taxes. At sustained volume, dedicated endpoints are often more economical — ask for a quote.

Engineered from silicon to API

Because Bastion runs the data center, the GPUs and the serving stack, every layer is tuned for throughput, latency and cost.

Autoscaling

Replicas follow traffic, with configurable floors and ceilings.

Cache-aware routing

Shared prefixes reuse KV cache for lower TTFT and cost.

Speculative decoding

Higher tokens per second on the Fast flavor.

Prompt caching

Repeated system prompts and documents billed at a discount.

Structured outputs

JSON mode and JSON-schema constrained decoding.

Tool calling

OpenAI-style tools for agents, with parallel calls.

Batch API

JSONL batches, 24-hour window, half-price tokens.

Observability

Per-key usage, latency percentiles and error rates.

Post-train on your data. Deploy in one click

Adapt open models to your domain, tone and tools — then serve them on the same infrastructure, with weights that stay yours.

LoRA fine-tuning

Low-cost adapters for Mistral, Phi, Granite and gpt-oss. Serve many adapters on one base.

Full fine-tuning

Full-parameter training on dedicated B300 clusters for maximum quality.

Preference tuning

DPO for alignment to your style and policies, with built-in eval runs.

Own your weights

Download checkpoints any time, or deploy them as private endpoints.

ServerlessDedicated endpointBatch
Best forPrototyping, variable trafficProduction, steady high volumeEvals, data generation, backfills
BillingPer tokenPer GPU-hour, committed termsPer token, −50%
LatencyLow, shared capacityLowest, reserved capacityWithin 24 hours
Custom / fine-tuned modelsLoRA adapters on supported basesAny supported open-weight modelCatalog models
NetworkingPublic HTTPSPublic or private (VPC peering)Public HTTPS
SLABest effort (preview)SLA-backed at general availabilityCompletion window

Backed by capacity we have secured

Token Factory runs only in Bastion-operated capacity, so you can choose exactly where your data is processed.

UNITED STATES

Inference served from Bastion-operated capacity in the United States.

580 MW and growing · U.S. data residency

EUROPEAN UNION

EU-only processing in Bastion-operated facilities in Europe for sovereign and regulated workloads.

300 MW and growing · data never leaves the EU

Works with the tools you already use

Any library that speaks the OpenAI API works with Token Factory.

O
OpenAI SDKOpenAI-compatible
L
LiteLLMOpenAI-compatible
L
LangChainOpenAI-compatible
L
LlamaIndexOpenAI-compatible
V
Vercel AI SDKOpenAI-compatible
H
HaystackOpenAI-compatible
C
CrewAIOpenAI-compatible
P
Pydantic AIOpenAI-compatible
A
AgnoOpenAI-compatible
D
DSPyOpenAI-compatible
O
Open WebUIOpenAI-compatible
n
n8nOpenAI-compatible
O
OpenAI SDKOpenAI-compatible
L
LiteLLMOpenAI-compatible
L
LangChainOpenAI-compatible
L
LlamaIndexOpenAI-compatible
V
Vercel AI SDKOpenAI-compatible
H
HaystackOpenAI-compatible
C
CrewAIOpenAI-compatible
P
Pydantic AIOpenAI-compatible
A
AgnoOpenAI-compatible
D
DSPyOpenAI-compatible
O
Open WebUIOpenAI-compatible
n
n8nOpenAI-compatible

Zero data retention

Prompts and completions are not stored after the request completes and are never used to train models.

U.S. or EU residency

Pin each project to U.S. processing or to EU-only processing.

Enterprise controls

Scoped API keys, per-project quotas, console SSO, audit logs and DPAs on request.

Join the Token Factory preview

Token Factory is in preview. Join the waitlist for early access and preview credits, or pre-order dedicated capacity. Tell us your models and volumes and we'll reach out as access opens.

By submitting, you agree to our privacy policy and terms of use. Access is subject to identity and restricted-party screening.

Questions, answered

Is Token Factory available today?

Token Factory is in preview. We are onboarding early-access customers from the waitlist and accepting pre-orders for dedicated endpoints. Prices and features shown are indicative until general availability.

Are all models open-weight?

Yes. Every model in the catalog has publicly released weights and a permissive license — Apache 2.0 or MIT — that allows commercial use and hosting. We deliberately list only permissively licensed models; models with restrictive or conditional licenses are not offered. The License column shows each one.

Is Token Factory really OpenAI-compatible?

Yes. Chat completions, completions, embeddings, images, audio transcription, batch and streaming follow the OpenAI API schema. Point your existing SDK at https://api.bastion-cloud.com/v1, use your Bastion API key and pick a model ID from the catalog.

Which models do you offer?

A focused catalog of open-weight models under permissive Apache 2.0 or MIT licenses, from OpenAI (gpt-oss, Whisper), Mistral AI, Microsoft (Phi-4, E5), IBM (Granite), Snowflake, Nomic AI and Black Forest Labs (FLUX.1 [schnell]). Other open-weight models can be deployed on dedicated endpoints on request, after license review.

What is the difference between Base and Fast?

Base is optimized for cost and suits most applications. Fast runs on reserved high-throughput capacity with speculative decoding for lower latency and more tokens per second, at twice the base price.

Do you store or train on my data?

No. By default, prompts and outputs are processed in memory and not retained after the request completes, and they are never used to train any model. Optional logging can be enabled per project for debugging.

Where is my data processed?

Only in Bastion-operated capacity. Each project can be pinned to the United States or to EU-only processing in our European facilities.

Who is responsible for export-control compliance?

Each customer is solely responsible for complying with U.S., EU and other applicable export-control and sanctions laws, including rules on advanced computing and AI compute. Bastion screens customers and may suspend access at any time, but that screening does not shift responsibility to Bastion. See Acceptable use & export compliance.

When should I move to a dedicated endpoint?

When you have steady, high-volume traffic, strict latency targets, need private networking or want to serve a custom model. Our team can model the break-even point against serverless pricing for your workload.

Build on open models. Keep control of what matters

Choose the models that fit your workload, deploy them the way your product needs and improve them with your own data — one path from first experiment to production, on capacity Bastion has secured.

Start with a token.
Scale to a cluster.