Hal 1.0
Radium's model for the hardest work in production: reasoning across sources, writing and fixing code, and agents that run for many steps.
Hal is the model Radium puts on the hardest work in production: reasoning across documents, writing and fixing code, and agents that keep calling tools until the job is done.
Radium selects open-weight models, tunes them and serves them behind one API, where a router and a reinforcement learning engine choose the model for each request.
The hard requests go to Hal.
Agents
Long loops of planning and parallel tool calls, with no tokens-per-minute limit to slow a job down.
Documents
Invoices checked against purchase orders, contracts compared clause by clause, and scans read as images.
Code
Writing, reviewing and fixing code across files, from the tools your team already runs, Claude Code included.
Analysis
Figures reconciled across reports, and answers that depend on a chain of documents.
4.6 times the output speed of Claude Opus.
Hal and Claude Opus ran the same 80 structured tasks with fixed answers, then 40 tasks at four concurrent requests for speed. Hal scored a little higher on accuracy and wrote its answers far faster, and it takes slightly longer to start. GPT-5.5 results on the same tasks are pending.
GPT-5.5 pending
GPT-5.5 pending
GPT-5.5 pending
GPT-5.5 pending
The infrastructure, measured by others.
One model ID, and the whole stack under it.
- PlatformThe Radium API, its router and RL engine, and a pool of tuned open-weight models.OpenAI- and Anthropic-compatible, so your SDK and prompts stay where they are.The router and RL engine choose the model for each request.Radium's model curation team selects and tunes the models.
- InfraDistributed computing across data center sites, on modified server firmware and Radium's own switch software.
- HardwareRacks of GPU servers in data centers across several sites.
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
- 128,000 tokens
- Input
- Text and images
- 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 Clarke and send the hard requests to Hal.
Hal
Reasoning across sources, 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.
Does Radium train its own foundation model?
Radium builds Hal from open-weight models that its model curation team selects and tunes, and serves them behind one API with a router and a reinforcement learning engine that choose the model for each request.
How do I move a workload to Hal?
Set the base URL to https://api.radium.cloud/v1, use a Radium API key and set the model to hal-1.0. Your SDK, prompts and tool schemas stay the same.
What does Hal cost?
$2.00 per million input tokens and $10.00 per million output tokens. New accounts get $25 of credit with no card needed.
Does Radium train on my data?
Your prompts and outputs are never used to train a model.
How was Hal benchmarked?
On Radium's own suite of 80 structured tasks with fixed answers and 40 tasks for speed, against Claude Opus, with GPT-5.5 results on the same suite pending. The full method, confidence intervals and limits are on the benchmark results page.
Put your hardest requests on Hal.
One line of code to switch.
A different class of performance.
Every new account gets $25 in credit. No card required.
