Hal 2.0
Radium's new flagship, built for long-horizon agents and complex software engineering. A 1M-token context window, text, image, video and audio in, and frontier-class results at a fraction of the cost.
Hal 2.0 is Radium's flagship: a trillion-parameter sparse mixture-of-experts model that goes head-to-head with the best frontier models on agentic coding and automation, at a fraction of the cost.
It activates 42B of its 1.02T parameters per token, so frontier capability costs a fraction as much to run, and it is served on Radium's throughput-optimized fleet behind the same API as Clarke and Tycho.
Work that runs long and has to come out right.
Agents
Terminal and computer-use agents that keep going: 89.9% on Terminal-Bench 2.1 and 82% on OSWorld-Verified, with a 1M-token context to finish long-horizon work.
Multimodal
Text, image, video and audio in one request, with up to 131K tokens of output.
Software engineering
Complex, multi-file engineering work, within 2 points of the top frontier models on DeepSWE v1.1.
Reasoning
Enterprise reasoning across a 1M-token context, for workloads where quality is non-negotiable.
Switching takes three values.
Point the SDK you already use at Radium with a new base URL, a Radium key and the model name. Switching back is the same change in reverse.
- Endpoints
- OpenAI- and Anthropic-compatible
- Context window
- 1M tokens
- Input
- Text, image, video and audio
- Structured output
- JSON mode and JSON schema
- Tool calling
- Parallel calls, strict schemas, multi-step loops
- Streaming
- OpenAI and Anthropic formats, with tool call deltas
- Rate limits
- 1,000 requests a minute, burst of 33, no tokens-per-minute limit
- Coming later
- Responses API, embeddings and batch
Start on Hal. Move work to Clarke and Tycho as it gets simpler.
Hal 1.0
Reasoning across documents, writing and fixing code, and long agent loops.
Clarke
RAG, copilots, tool calling and most production traffic.
Tycho
Classification, extraction and routing at volume, on text.
Questions about Hal 2.0.
What is new in Hal 2.0?
Hal 2.0 is the next generation of Hal. It reads up to 1M tokens of context, up from 128K on Hal 1.0, and writes up to 131,000 tokens in a single response, up from 32,768. It takes text, image, video and audio as input and sits in the same class as Anthropic's Opus 5.5 and OpenAI's GPT-6.1 Sol.
How do I get access during the preview?
Sign in to your Radium account and request access from the dashboard. Capacity is limited while Hal 2.0 is in preview, so access is granted account by account, and the model string appears in your account once yours is approved. New accounts can request it as soon as they sign up.
Is Hal 2.0 replacing Hal 1.0?
Yes. When Hal 2.0 leaves preview, Hal 1.0 is retired and requests sent to hal-1.0 are served by Hal 2.0, so existing integrations move over without a code change. Until then, Hal 1.0 stays available to every account.
Does Hal 2.0 cost more than Hal 1.0?
No. Hal 2.0 is priced the same as Hal 1.0, at $2.00 per million input tokens and $10.00 per million output tokens, with cached input at 10% of the input price. Rate limits are the same too: 1,000 requests per minute with no tokens-per-minute cap.
Do I need to change my code?
Only the model name. With the OpenAI SDK, set the base URL to https://api.radium.cloud/v1, and with the Anthropic SDK, set it to https://api.radium.cloud. Use your Radium API key and the Hal 2.0 model string shown in your account once access is approved. Your prompts and tool schemas stay the same.
When does Hal 2.0 leave preview?
Radium will announce general availability on this page and by email to every account, along with the date Hal 1.0 is retired.
Give Hal 2.0 the job your current model keeps getting wrong.
Hal 2.0 is in early access.
Request it in your Radium account.
Every new account gets $25 in credit. No card required.
