Private beta · Waitlist

Frontier capability, unrestricted.

A frontier-class model that completes the work mainstream APIs refuse. Built on GLM-5.2.

Coming soonKimi K3
Current monthly spend on LLM APIs

We convert this to capacity. A rough guess is fine.

Who will use it?
What will you build?

Pick all that apply.

We'll send a confirmation now and one email when your spot opens. No spam.

Early access members get their first top-up matched, dollar for dollar. Add $50, get $50.

Evaluation preview

Capability preserved, in the numbers.

Dionysus remains close to or improves on its GLM-5.2 FP8 baseline across science, instruction following, software engineering, agents, and telecom.

GLM-5.2 FP8Dionysus

Reported 0–100 scores. Each track is zoomed to a ±2-point window around the pair so small movements read clearly; higher is better. Exact values and signed deltas sit alongside every row.

Context window

Full context size

1Mtokens

Movement vs GLM-5.2 FP8 baselineavg Δ +0.20 across 6 evals

GPQA Diamond

Graduate science

89.52 89.24Δ -0.28

SciCode

Scientific coding

49.85 48.87Δ -0.98

IFBench

Instruction following

74.95 75.61Δ +0.66

τ²-Bench Telecom

Telecom agents

97.90 98.50Δ +0.60

Terminal-Bench 2.1

Agentic terminal

81.00 82.40Δ +1.40

SWE-bench Pro

Software engineering

62.10 61.90Δ -0.20

More benchmarks, full technical reports, evaluation methodology, and limitations will be published at launch.

Capability

Frontier-class, where it counts.

Capability you can actually use, measured against frontier general-purpose models.

01

Fewer artificial refusals

Where mainstream APIs return a safety template, Dionysus-1 returns a completion. Built for legitimate work in domains general-purpose models routinely over-refuse.

02

Frontier-class base

Built on GLM-5.2, with capability preserved and measured: full benchmark suite, KL-divergence checks, regression tests, and human evals in the technical report.

03

Private by default

No training on your prompts. No retention mode. OpenAI-compatible API with streaming and prompt caching, so it drops into what you already build with.

What happens next

A clear path from signup to access.

01

Confirmation now

We save your place and send a transactional confirmation to the address you provide.

02

Capacity-based invitation

We review demand and invite members in batches as private beta capacity opens.

03

$50 matched at launch

Early access members get their first top-up matched dollar for dollar. Add $50, get $50.

Contact

Have something specific in mind?

Tell us about model access, enterprise requirements, partnerships, or security and compliance needs. This form is separate from joining the waitlist.

Replies are reviewed by the Wallachia Labs team.

Your message is used only to review and respond to this inquiry. See our privacy notice.

FAQ

Before you join.

The practical details we can answer before private beta capacity opens.

Who is the private beta for?

The beta is intended for individual builders, research teams, security practitioners, and companies that need an OpenAI-compatible model with fewer artificial refusals.

When will invitations be sent?

Invitations will be released in batches as inference capacity opens. Joining earlier helps us understand demand, but it does not guarantee a specific access date.

What does OpenAI-compatible mean?

The planned API follows the familiar OpenAI client and chat-completions shape, so existing integrations can switch by changing their base URL and API key.

How does the first top-up match work?

Eligible early access members will receive a dollar-for-dollar match on their first top-up, up to $50. Add $50 and receive $50 in matching credit.

What happens to the information I submit?

Waitlist information is used to manage access and plan capacity. Contact-form information is used only to review and respond to the inquiry. The privacy page explains the details.