The work, dated.

Newcastle Compute v1.0 — soft launch

Today we open Newcastle Compute to the first cohort of users. Initial capacity at Leading Edge Data Centres, Mayfield West: 8× H200, 16× H100, 16× RTX PRO 6000 Blackwell. B200 and AMD MI325X capacity available on request with short lead times.

Hosted inference is live for nine open-weights models behind a single OpenAI-compatible API. Fine-tuning service accepts jobs immediately. Access is invitation-based while we shake out the self-serve flow; tell us what you're building for a same-day API key.

Qwen-3 VL deployed

Vision-language variant of Qwen-3 live on H100. Document understanding, chart and table reading, screenshot QA, video frame analysis. Image input via URL or base64. Available as qwen-3-vl.

Qwen-3 235B Instruct deployed

Frontier Qwen-3 MoE chat model live on H200. Multi-node tensor-parallel deployment via vLLM 0.7.2. Initial benchmarks: 142 tokens/sec generation, 0.36s first-token latency at batch 1. Available as qwen-3-235b-instruct.

RTX PRO 6000 Blackwell capacity online

16× RTX PRO 6000 Blackwell Server Edition cards online for inference workloads. 96GB GDDR7 each, FP4 support. Replaces L40S as the inference workhorse — measured throughput uplift of ~5.4× on Qwen-3 14B vs prior L40S benchmarks.

DeepSeek-V3.1 added

671B parameter MoE serving from a multi-node H200 cluster. Strong reasoning and code performance. Available as deepseek-v3.1. Already a customer favourite for agentic workflows.

vLLM 0.7.2 across all inference endpoints

Rolled out vLLM 0.7.2 with improved speculative decoding and FP8 KV cache. Observed throughput improvement of 18–24% across the Qwen-3 family and 12% on Llama 3.3.

Mayfield West cluster commissioned

First production GPU capacity online at Leading Edge Data Centres, Mayfield West. Direct-to-chip liquid cooling on H200/B200 trays; rear-door heat exchangers elsewhere. Two carrier-diverse fibre paths. 50 kW+ racks operational.

Qwen-3 family + Llama 3.3 + Mistral Large 2 brought up

Initial model lineup deployed for pre-launch testing: Qwen-3 (235B, 72B, 14B, Coder variants), Llama 3.3 70B Instruct, Mistral Large 2, Qwen-3 Embed (large + base). All running on H200/H100/RTX PRO 6000 backed endpoints.