# Doubleword - llms.txt
# https://doubleword.ai/llms.txt
# Last updated: 2026-03-23

## 1. Company Overview

Doubleword is a London-based AI inference company that provides dramatically cheaper LLM inference at scale. The core product is a hosted batch inference cloud for high-volume, non-real-time AI workloads.

**Who it's for**: Engineering teams running large-scale AI workloads - background agents, synthetic data pipelines, document processing, model evaluations, embeddings - where latency is not the primary constraint but cost and throughput are.

**Core value proposition**: Run massive, asynchronous LLM workloads at wholesale compute prices. Up to 97% cheaper than real-time APIs and up to 90% cheaper than competing batch APIs (e.g. Anthropic Batch, OpenAI Batch). Same intelligence class, verified by independent benchmarks from Artificial Analysis.

**Tagline**: AI Inference, Built for Scale.

**Headquarters**: London, UK
**Contact**: hello@doubleword.ai
**Website**: https://doubleword.ai

---

## 2. Product

### Three Inference Modes

Doubleword offers three delivery tiers, all accessible through a single OpenAI-compatible API:

**Dev Mode (Real-time)**
- Full-price, zero-wait inference
- For iterating on prompts and models during development
- Standard streaming responses

**Priority (~1 hour SLA)**
- Up to 50% cheaper than real-time
- Guaranteed 1-hour completion window
- Best for background agents that need results fast
- completion_window="priority"

**Overnight (~24 hours SLA)**
- Up to 75% cheaper than real-time
- Deepest discounts for big batch jobs
- completion_window="24h"

### The Autobatcher

The Autobatcher is a drop-in replacement for standard async OpenAI calls. Install with `pip install autobatcher`. It transparently collects and batches requests over a configurable window, capturing batch-tier pricing with zero code refactoring. No .jsonl files, no polling logic, no manual parsing.

### API

- **Base URL**: https://api.doubleword.ai/v1/
- **Compatibility**: Fully OpenAI-compatible. Supports chat completions, embeddings, JSON-mode, tool calling, function calling, structured outputs, and the /responses endpoint.
- **Authentication**: API key via the Doubleword console at https://app.doubleword.ai
- **Batch endpoint**: Uses the standard OpenAI batches.create() method with custom completion_window values.

### Key Features

- Results stream as they complete - no waiting for the whole batch to finish
- Mixed request routing across chat completions, embeddings, and JSON-mode in a single batch window
- Webhooks to plug directly into existing pipelines and event-driven architectures
- Automatic retries and failover
- Guaranteed delivery - if the SLA is missed, the job is free
- Cancel or retry any batch at any time
- No credit card required, no minimum spend, pay only for tokens used

---

## 3. Use Cases

Doubleword publishes runnable "Workbooks" - real examples with real costs. Each links to GitHub code.

| Use Case | Description | Example Cost | Key Insight |
|---|---|---|---|
| Async Agents | Deep research with recursive agent trees | $0.34 for 47 agents | 1-hour SLA enables multi-round agentic workflows |
| Synthetic Data Generation | Generate training data with quality filtering | $3.21 for 10K samples | 3-stage pipeline in 3 hours, not 3 days |
| Data Processing Pipelines | Clean and enrich messy records | $0.80 for 50K records | LLM-powered normalization at pipeline scale |
| Embeddings | Semantic search over document corpus | $0.03 for 1.6M tokens | 70% cheaper than OpenAI for same quality |
| Model Evals | Benchmark models on GSM8K | $0.21 for 1,319 questions | Comprehensive evaluation becomes routine |
| Bug Detection Ensemble | Classify security vulnerabilities | $0.40 for 4,642 samples | Run twice for calibration, still under $1 |
| Dataset Compilation | Build company datasets via search + LLM | $1.05 for 188 companies | 100% recall vs Gartner Magic Quadrant |
| Structured Extraction | Extract fields from scanned receipts | $0.12 for 626 receipts | Qwen3-30B beats GPT-5.2 at 1/12th cost |
| Image Summarization | Caption images for social media | $0.10 for 1,000 images | Vision batch makes captioning automatic |
| Asynchronous Event Listeners | Process events asynchronously at scale | $0.15 for 10K events | Decouple event handling from request paths |
| ETL & Pipeline Sanitization | Sanitize and transform data in ETL pipelines | $0.50 for 100K records | LLM-powered sanitization catches what regex misses |

**Not suited for**: Chat UIs, latency-critical paths. Doubleword optimizes for cost and throughput, not latency.

---

## 4. Models

All models are open-weight and accessible through the OpenAI-compatible API. New models can be added within 24 hours on request.

### Text Models

| Model | Model ID | Context | Benchmark Score (AA) |
|---|---|---|---|
| Qwen3.5-397B-A17B | Qwen/Qwen3.5-397B-A17B | 256k | 45 |
| Qwen3.5-35B-A3B-FP8 | Qwen/Qwen3.5-35B-A3B-FP8 | 256k | 37 |
| Qwen3.5-9B | Qwen/Qwen3.5-9B | 256k | 32 |
| Qwen3.5-4B | Qwen/Qwen3.5-4B | 262k | 27 |
| GPT-OSS-20B | openai/gpt-oss-20b | 128k | 25 |
| Nemotron-3-Super-120B-A12B | nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4 | 256k | 36 |
| Qwen3-14B-FP8 | Qwen/Qwen3-14B-FP8 | 32k | 13 |

### Vision Models

| Model | Model ID | Context | Benchmark Score (AA) |
|---|---|---|---|
| Qwen3-VL-235B-A22B | Qwen/Qwen3-VL-235B-A22B-Instruct-FP8 | 256k | 21 |
| Qwen3-VL-30B-A3B | Qwen/Qwen3-VL-30B-A3B-Instruct-FP8 | 256k | 16 |

### Embedding Models

| Model | Model ID | Context | Benchmark Score |
|---|---|---|---|
| Qwen3-Embedding-8B | Qwen/Qwen3-Embedding-8B | 32k | MTEB 70.6 |

---

## 5. Pricing

All prices are per 1M tokens. No minimum spend. No credit card required. Pay only for tokens used.

### Qwen3.5-397B-A17B
- Priority: $0.30 input / $1.80 output
- Overnight: $0.15 input / $1.20 output

### Qwen3.5-35B-A3B-FP8
- Priority: $0.07 input / $0.30 output
- Overnight: $0.05 input / $0.20 output

### Qwen3.5-4B
- Priority: $0.05 input / $0.08 output
- Overnight: $0.04 input / $0.06 output

### Qwen3.5-9B
- Priority: $0.04 input / $0.35 output
- Overnight: $0.03 input / $0.29 output

### GPT-OSS-20B
- Priority: $0.03 input / $0.20 output
- Overnight: $0.02 input / $0.15 output

### Nemotron-3-Super-120B-A12B
- Priority: $0.23 input / $0.56 output
- Overnight: Currently promotional / free

### Qwen3-VL-235B-A22B
- Priority: $0.15 input / $0.55 output
- Overnight: $0.10 input / $0.40 output

### Qwen3-VL-30B-A3B
- Priority: $0.07 input / $0.30 output
- Overnight: $0.05 input / $0.20 output

### Qwen3-14B-FP8
- Priority: $0.03 input / $0.30 output
- Overnight: $0.02 input / $0.20 output

### Qwen3-Embedding-8B
- Priority: $0.03 input (no output cost)
- Overnight: $0.02 input (no output cost)

### Competitor Comparison (Qwen3.5-397B class, output per 1M tokens)

- Anthropic Claude Opus 4.5 Realtime: $25.00
- Anthropic Claude Opus 4.5 Batch (24H): $12.50
- OpenAI GPT-5.2 Realtime: $14.00
- OpenAI GPT-5.2 Batch (24H): $7.00
- Doubleword Qwen3.5-397B Priority (~1h): $1.80 (up to 86% less)
- Doubleword Qwen3.5-397B Overnight (24h): $1.20 (up to 90% less)

---

## 6. Why Doubleword

### vs Real-Time APIs (for agents)

Real-time APIs are built for chatbots, not agents. Most agent tasks - research, analysis, content generation - don't need millisecond latency. Doubleword provides:
- Up to 95% cheaper per token
- Built-in retries and failover (no DIY retry logic)
- Managed throughput at scale (no rate limits throttling agents)
- Full tool call support, function calling, structured generation
- One-line migration: swap your base URL, keep your code

### vs Traditional Batch APIs

Traditional batch APIs (OpenAI Batch, Anthropic Batch) save money but force complexity:
- **Getting started**: They require .jsonl boilerplate and file uploads. Doubleword: `pip install autobatcher` - done.
- **Cost savings**: They offer ~50% off. Doubleword: up to 97% off.
- **Speed**: They take 24 hours and may expire. Doubleword: Priority in ~1hr, or 24hr for lowest cost.
- **Failure handling**: They have silent failures and expired batches. Doubleword: guaranteed delivery or your money back.
- **Getting results**: They require waiting for everything to finish. Doubleword: streamed back as they complete.
- **Integration**: They require rewriting your app. Doubleword: webhooks, streaming, OpenAI-compatible.

### Inference Stack

Doubleword optimizes every layer of the inference stack for throughput and cost:

1. **Gateway** - Open-source Rust-based model gateway (Control Layer) with 450× less overhead than LiteLLM. Handles routing, access controls, logging, monitoring.
2. **Scheduling & Orchestration** - Priority-based queue management, demand smoothing, load balancing, autoscaling.
3. **Runtime Engine** - Built on TensorRT-LLM, SGLang, vLLM, Dynamo with proprietary throughput optimizations: speculative decoding (QueueSpec, Spacelike), continuous batching, memory-efficient attention, KV-cache optimization, ZeroDP.
4. **Model** - Aggressive quantization (INT8, FP8, INT4), KV cache quantization, LoRA adapter fusion, custom kernel implementations.
5. **Hardware** - Multi-provider cost optimization, right accelerator per workload, disaggregated inference, strategic spot instance usage.

---

## 7. How It Works

### Batch Workflow

1. Create an API key at https://app.doubleword.ai
2. Install the OpenAI Python SDK or use the Autobatcher: `pip install autobatcher`
3. Set base_url to https://api.doubleword.ai/v1/
4. Submit batch jobs using client.batches.create() with your chosen completion_window
5. Results stream back as they complete - process outputs immediately
6. Use webhooks for event-driven pipeline integration

### Code Example (Priority Batch)

```python
from openai import OpenAI

client = OpenAI(
    api_key="your-api-key",
    base_url="https://api.doubleword.ai/v1/"
)

result = client.batches.create(
    input_file_id=batch_input_file_id,
    endpoint="/v1/chat/completions",
    completion_window="priority",
    metadata={"description": "Structure Extraction"}
)
```

### Code Example (Real-time Dev Mode)

```python
from openai import OpenAI

client = OpenAI(
    api_key="your-api-key",
    base_url="https://api.doubleword.ai/v1/"
)

result = client.chat.completions.create(
    model="Qwen/Qwen3-VL-235B-A22B-Instruct-FP8",
    messages=[{"role": "user", "content": "Hello!"}]
)
```

---

## 8. Getting Started

1. **Sign up**: Visit https://app.doubleword.ai - no credit card required
2. **Get your API key**: Available in the console immediately after sign-up
3. **Free tier**: Get started for free
4. **Quickstart**: Read the docs at https://docs.doubleword.ai/batches/skill#installation
5. **Async Pipeline Builder**: Interactive onboarding at https://onboarding.doubleword.ai/

---

## 9. About

Doubleword is a London-based team of physicists and systems engineers obsessed with inference efficiency. They have built their own inference engines, published research, optimized kernels in CUDA and Triton, and deployed production infrastructure inside regulated enterprises.

**Mission**: What becomes possible when inference is 100x cheaper?

**Team**: Engineers at the intersection of systems engineering and ML infrastructure.

**Hiring**: Doubleword is actively hiring. Open roles at https://doubleword.ai/careers

---

## 10. Links

- **Website**: https://doubleword.ai
- **Console / Sign Up**: https://app.doubleword.ai
- **Documentation**: https://docs.doubleword.ai
- **Batch Docs**: https://docs.doubleword.ai/batches
- **Technical Blog**: https://blog.doubleword.ai
- **GitHub**: https://github.com/doublewordai
- **LinkedIn**: https://www.linkedin.com/company/86740135
- **Async Pipeline Builder**: https://onboarding.doubleword.ai/
- **Resource Centre**: https://www.doubleword.ai/resources
- **Pricing**: https://doubleword.ai/pricing
- **Models**: https://doubleword.ai/models
- **Use Cases / Workbooks**: https://doubleword.ai/use-cases
- **Savings Calculator**: https://doubleword.ai/calculator
- **Careers**: https://doubleword.ai/careers
- **About**: https://doubleword.ai/about
- **Contact**: hello@doubleword.ai
- **Privacy Policy**: https://doubleword.ai/privacy
- **Terms of Service**: https://doubleword.ai/terms-of-service
- **Data Usage Policy**: https://doubleword.ai/data-usage-policy

---

## FAQ

**Why is it cheaper?**
Doubleword built an inference stack optimized for high throughput and low cost from the ground up. By optimizing at every layer - hardware, runtime, and orchestration - they achieve significantly better unit economics than providers who bolt batch features onto real-time infrastructure.

**What happens if the SLA is missed?**
Guaranteed delivery. Unlike providers that may expire batches, Doubleword commits to the chosen SLA. If it's missed, the request won't be expired and won't be charged.

**Can I get results before the batch completes?**
Yes. Results are streamed as they're processed - no waiting for every request to complete.

**Is it secure?**
Yes. Full control of data with deletion on demand. DPA available. SOC2 and ISO27001 in progress. Detailed security policies available on request.

**What if a run fails?**
Automatic retries and recovery. Partial results are saved; jobs can resume from where they left off.

**Which models are supported?**
Popular open-weight LLMs, vision models, and embedding models. New models added within 24 hours on request.

**Should I use this for my chat app?**
No. Doubleword optimizes for cost and throughput, not latency. For latency-critical chat, use a real-time provider.

---

## Pages

- [Home](https://doubleword.ai/): High throughput async AI inference at scale.
- [Pricing](https://doubleword.ai/pricing): Per-model pricing for Async and Overnight tiers.
- [Models](https://doubleword.ai/models): Supported open-weight text, vision, and embedding models.
- [Use Cases](https://doubleword.ai/use-cases): Runnable workbooks with real costs.
- [Calculator](https://doubleword.ai/calculator): Estimate savings vs real-time and batch APIs.
- [Inference Stack](https://doubleword.ai/inference-stack): How the Doubleword stack is optimized for throughput.
- [Why Doubleword](https://doubleword.ai/why-doubleword): Comparison vs real-time and traditional batch APIs.
- [Why Batch Inference](https://doubleword.ai/why-batch-inference): When async batch beats real-time inference.
- [Why Agents](https://doubleword.ai/why-agents): Why agent workloads belong on async inference.
- [Compare](https://doubleword.ai/compare): Side-by-side comparison with other inference providers.
- [Customer Stories](https://doubleword.ai/customer-stories): How teams use Doubleword in production.
- [Glossary](https://doubleword.ai/glossary): Definitions of inference and batch terminology.
- [About](https://doubleword.ai/about): The team and mission behind Doubleword.
- [Careers](https://doubleword.ai/careers): Open roles at Doubleword.
- [Privacy Policy](https://doubleword.ai/privacy): How we handle data and privacy.
- [Terms of Service](https://doubleword.ai/terms-of-service): Terms governing use of Doubleword.
- [Data Usage Policy](https://doubleword.ai/data-usage-policy): How customer data is used and retained.