Available Models

Frontier open-weight models, fully managed.

Every model on this page ships as open weights under an open-source licence. They run on your dedicated in-region node — we manage the serving stack, updates, and tuning, and your token data never leaves the hardware.

Start an Enterprise Workspace arrow_forward See node pricing

The lineup

Three open-weight families cover the spectrum from agentic multimodal coding to the hardest long-horizon agentic work.

Agentic All-Rounder

MiniMax M3

MiniMax · MoE · 1M context

59.0SWE-bench Pro

Frontier coding combined with native multimodal input and computer use — the first open-weight model built for 24/7 autonomous agents that browse, click, and code.

  • check1M-token context window
  • checkNative image, video, and desktop operation
  • checkStrong on browsing benchmarks (BrowseComp 83.5%)
Open-Weight Leader

GLM-5.2

Z.AI · ~750B MoE · MIT license

62.1SWE-bench Pro

The strongest open-weight coder available today. Deep reasoning and tool use across full repositories, with a permissive MIT licence and Anthropic-compatible API surface.

  • checkBest open-weight SWE-bench Pro score
  • checkTerminal-Bench 2.1: 81.0%
  • checkBeats GPT-5.5 on real-world coding
Code Specialist

Kimi K2.7 Code

Moonshot AI · ~1T total · 32B active

58.6SWE-bench Pro (K2.6)

A coding-focused refinement of the trillion-parameter K2 family, tuned for the hardest, longest-horizon agentic tasks — with ~30% fewer reasoning tokens than K2.6.

  • checkBuilt for long-horizon agentic coding
  • check~30% fewer thinking tokens than K2.6
  • checkLaunch benchmarks are vendor-internal; K2.6 family score shown

Parameter counts and scores are vendor-reported for the latest released checkpoint per family, June 2026. We roll out new checkpoints to managed nodes as they are released — your endpoint stays the same.

The Unlock · SWE-bench Pro 2026

Open weights are closing on the frontier.

In 2026 the open-weight pack reached within ~2 points of Claude Opus 4.7 on SWE-bench Pro, the contamination-resistant real-world coding benchmark — and every one of them runs on hardware you control.

SWE-bench Pro · Resolved % Claude Opus 4.7 = frontier reference · 64.3
GLM-5.2Z.AI
62.1
MiniMax M3MiniMax
59.0
Kimi K2.7 CodeMoonshot AI
58.6

K2.6 family score — K2.7 Code launched with vendor-internal benchmarks only.

DeepSeek V4DeepSeek
55.4

Source: SWE-bench Pro (Pass@1), vendor-reported 2026 results, latest released checkpoint per model family. Claude Opus 4.7 (64.3) shown as the closed-frontier reference; standardized cross-model scaffolds run lower for every model. The gap is small and closing fast — near-frontier coding quality on infrastructure you own is exactly what makes a sovereign coding server viable today.

Open weights on your node vs. a frontier API

The last few benchmark points buy you a foreign vendor in your data path. Here is what you trade — and what you keep.

SWE-bench Pro (best available)

Frontier API (e.g. Claude Opus 4.7)
64.3
Open weights on your syndicAI node
62.1 (GLM-5.2)

Where your source code goes

Frontier API (e.g. Claude Opus 4.7)
US-owned cloud, on every request
Open weights on your syndicAI node
Never leaves your node

Jurisdiction

Frontier API (e.g. Claude Opus 4.7)
US CLOUD Act applies
Open weights on your syndicAI node
Swiss or EU region, by design

Cost model

Frontier API (e.g. Claude Opus 4.7)
Per token — ~CHF 5K/developer/month at agentic volume
Open weights on your syndicAI node
Flat monthly per node, unlimited tokens

Rate limits

Frontier API (e.g. Claude Opus 4.7)
Throttled at peak load
Open weights on your syndicAI node
None — the hardware is yours

Model weights

Frontier API (e.g. Claude Opus 4.7)
Closed, vendor-controlled
Open weights on your syndicAI node
Open, inspectable, pinned to your node

Fully managed, on your hardware

Running a frontier open-weight model well is an infrastructure problem, and it is the part we own:

  • Serving stack — vLLM with tensor parallelism across 1–4 GPUs (96 GB each), tuned per model for context length and concurrency.
  • Updates — new checkpoints are rolled out to your node as vendors release them. Your endpoint and keys never change.
  • Sizing — pick the model; we size and price the hardware. The model-fit estimator shows which node sizes host which models.
  • The largest flagships — very large configurations (for example GLM-5.2 or Kimi K2.7 Code at maximum context) can exceed a standard 4-GPU node. Get in touch and we will scope a custom configuration.

Model FAQ

Pick your model. We size and price the hardware.

See which node sizes host which models, with flat monthly prices — or start a workspace and provision your first node.

View enterprise pricing arrow_forward Start an Enterprise Workspace