Changelog

GPU availability, new AI models, pricing updates, and platform releases.

RTX 5090 Now Available — 32GB GDDR7

NVIDIA RTX 5090 pods are now available on the Lium network (Bittensor SN51).

  • 32GB GDDR7 VRAM — fits larger models than RTX 4090
  • Higher memory bandwidth: 1.8 TB/s vs 1.0 TB/s on RTX 4090
  • Per-second billing, deploy in under 60 seconds

Confidential Agent Pilot Program — 30 Days Free

Law firms, accounting firms, and financial analysts can now apply for a free 30-day pilot of the Confidential Agent Platform.

  • Zero cost, no setup fee, no credit card required
  • Full access to all 8 agent templates with enterprise tools
  • Includes dedicated onboarding and support
  • Intel TDX encrypted enclaves — documents never leave the hardware boundary

Kimi K2.5 Added to Inference API

Kimi K2.5 (MoonshotAI) is now available via the OpenAI-compatible inference API.

  • 1T parameter MoE model — long-context reasoning
  • Strong performance on coding and agentic tasks
  • Compatible with all OpenAI SDK clients

Qwen3-235B-TEE Available in Confidential Compute

The Pro tier of the Confidential Agent Platform now uses Qwen3-235B (235B MoE parameters) running inside Intel TDX enclaves.

  • 7x more powerful than the Starter 32B model
  • 262K context window — ingest entire contracts or financial reports in one pass
  • Same hardware security: Intel TDX + NVIDIA Protected PCIe

B200 192GB Now Available

NVIDIA B200 (Blackwell) 192GB HBM3e pods are now available on the Lium network.

  • 192GB HBM3e VRAM — runs 70B+ models at full precision without quantization
  • NVLink 5.0 for multi-GPU configurations up to 8x B200
  • Available via Browse Pods and API

IDOR Security Fixes — Admin Auth Hardened

Two security issues were identified and fixed: IDOR vulnerability in pod API endpoints and admin authentication bypass.

  • All pod/volume endpoints now verify ownership against authenticated userId
  • Admin routes require explicit role check — no more implicit elevation
  • Internal security audit completed — no customer data was accessed

DeepSeek-R1-TEE Available — Enterprise Tier

DeepSeek-R1 (685B reasoning model) is now available in the Enterprise tier of the Confidential Agent Platform.

  • Chain-of-thought reasoning for multi-step legal and financial analysis
  • CFA-level financial modeling, IACCM-level contract review
  • Runs inside Intel TDX enclave — same hardware security as all other tiers
  • 163K context window

Fine-Tuning: GRPO Support Added

GRPO (Group Relative Policy Optimization) is now supported as a training method for fine-tuning jobs.

  • Reinforcement learning with custom Python reward functions
  • Supports any reward signal: rule-based, model-based, or custom
  • Available on Gradients network (Bittensor SN56)

OpenAI SDK Compatibility Verified — LangChain, CrewAI, OpenClaw

Compatibility verified and documented for major agent frameworks using the VoltageGPU inference API.

  • LangChain: tested with ChatOpenAI(base_url=...)
  • CrewAI: tested with all agent loop patterns
  • OpenClaw: full compatibility with tool-calling endpoints
  • One-line migration: change base_url to https://api.voltagegpu.com/v1

H200 141GB — Expanded Capacity

H200 pool expanded from 40 to 60+ available instances on the Lium network.

  • 141GB HBM3e VRAM — fits large models at full precision
  • Used by Bittensor validators for top-tier inference benchmarks

Confidential Agent Platform — Public Beta

8 AI agent templates for legal, finance, healthcare, and compliance launched in public beta with Intel TDX hardware encryption.

  • Contract Analyst, Financial Analyst, Compliance Officer, Medical Analyst
  • Due Diligence, Cybersecurity Analyst, HR Analyst, Tax Analyst
  • Free tier: 5 analyses/day, no signup required
  • Hardware: Intel TDX + NVIDIA H200 Protected PCIe

Per-Second Billing — All GPU Pods

Billing granularity improved from per-minute to per-second for all GPU pods.

  • Stop any pod instantly — pay only for exact compute time used
  • No minimum billing period
  • Significant cost savings for short workloads (inference runs, quick tests)

Platform Launch — GPU Compute + AI Inference

VoltageGPU launched with GPU Compute (Lium SN51) and AI Inference API (Chutes SN64).

  • RTX 3090 to H200 GPUs available on-demand
  • 66+ AI models via OpenAI-compatible API
  • Per-second billing from day one
  • SIREN 943 808 824 — EU-registered, GDPR Art. 25 compliant