LLM inference and training, on-premises or hybrid, explained
Short, verified analyses on on-premises AI platforms, open-weight models, LLM costs, agentic AI, and French and European compliance.
Local AI platform and open-source LLM: the analysis of an on-premise choice
A local AI platform on an open-source LLM: confidentiality, permanence and control of the model secured, licence and threshold manageable. Benchmarks, licences and costs verified.
DeepSeek V4 Flash Vision Exp: DeepSeek's first multimodal model
DeepSeek released its first open-weight multimodal model on 31 August 2026: what the vision tower adds, and what the experimental label commits you to.
Open weight models for local deployment 2026: complete comparison
The open-weight models that run on-premises in 2026: real size, active parameters, context window and the hardware each one needs, tier by tier.
Costed scenario: DeepSeek V4 on DGX Station for an 80-person SME
A sizing scenario, not a customer story: DeepSeek V4 Flash 0731 on a DGX Station GB300 for an 80-person SME. Estimated throughput, 3-year TCO, break-even against APIs.
DeepSeek V4 Flash 0731 on DGX Station and 2× DGX Spark cluster
DeepSeek V4 Flash 0731 (284B, 13B active, 1M context) on DGX Spark, a 2× Spark cluster and DGX Station GB300: vLLM 0.25+ recipes, benchmarks and real cost.
Published benchmark results of DeepSeek V4 Flash 0731 (open weight, 2026)
Vendor-published benchmark results for DeepSeek V4 Flash 0731 and GLM 5.2 (open weight, 2026): Terminal Bench, NL2Repo, GPQA Diamond, HLE. No estimates.
Independent benchmark: DeepSeek V4 Flash 0731 vs Claude Opus 5 vs GPT-5.5
DeepSeek V4 Flash 0731 (284B, 13B active, 1M context): vendor-published scores, real cost per token against the APIs, and what running on-premises changes.
Local LLM: which open-source AI model to choose?
Comparison of the best open-source LLMs to run locally in 2026: GLM, DeepSeek, Kimi, Mistral, Qwen. Size, strengths, licence and NVIDIA, HPE, Dell hardware.
Which server for AI? Guide and prices by range
How much does an AI server cost in 2026? From €6,100 ($6,956) for a DGX Spark workstation to about €451,000 excl. VAT for an 8-GPU server, tier by tier.
On-premises or API: from how many users does a local LLM pay off?
The LLM cost break-even point, the role of the KV cache, and the comparison with pay-per-token pricing.
GLM 5.3 deep on-premise: which hardware, at what speed?
A full study of GLM 5.3: 744 billion parameters, 40 active, a one-million-token context, and measured throughput from DGX Spark to the GB300 NVL72 rack.
Mistral Large 4 "le Chonk": what is verified, and which hardware can serve it
1.05 T parameters, 49 B active, weights expected late October. Cross-checked facts, contradictions, independent measurements, FP8 and NVFP4 memory, machines.
Full Stack AI Company: Building with On-Premises AI, Hugging Face and NVIDIA
Architecture to build a sovereign full stack AI enterprise: GPU hardware, NVIDIA AI Enterprise, Hugging Face models, agentic swarms via MCP and predictive TCO.
ASRock AI BOX-A395 + Qwen3.8-27B: 256K context on Strix Halo 128 GB
The ASRock AI BOX-A395 (Ryzen AI Max+ 395, 128 GB shared LPDDR5X) runs multimodal Qwen3.8-27B at its native 262,144-token context, using 30% of memory.
Running AI locally on your PC
Two ways to run AI on your own machine in 2026: free with open-weight models, or a DGX Spark mini-supercomputer (ASUS, Dell, Lenovo, MSI).
From your GPU server to a production LLM service
A seven-phase method for moving from a GPU server to an LLM service in production, with a deliverable and an exit criterion for each phase.
Tutorial: Operating a Digital Organisational Twin (DOTS)
Step-by-step visual tutorial: operating an on-premises and sovereign Digital Organisational Twin, monitoring 26 agent roles, arbitrating HTTP 409 quarantines and enforcing quotas.
On-Premises LLM Training and RFT: From QLoRA to Production Serving
Complete method to specialise a 27-billion parameter model on site: reinforcement curation, QLoRA adaptation, loss masking and vLLM production serving.
Advanced enterprise RAG patterns: what separates a demo from a system that holds
Chunk with context, filter by clearance before scoring, decompose, return wide, measure recall: the five patterns that make a RAG system hold in production.
AI orchestration engine: routing every request to the right model
Classify the request, pick the model, pass through a gateway with quotas: how an orchestration engine decides what stays on-premises and what leaves for an API.
Sensitive data and AI: what a zero-retention policy really guarantees
Zero retention is a contractual commitment, not an observable property. The four places data settles anyway, and what running on-premises changes.
From a chatbot to an agentic platform
The harness, three-layer memory and skills turn a chatbot into an agentic platform that executes tasks and capitalises on procedures.
LLM inference engines compared in 2026: vLLM, SGLang, TensorRT-LLM, llama.cpp and Dynamo
2026 comparison of LLM inference engines on Blackwell: vLLM, SGLang, TensorRT-LLM, Dynamo, llama.cpp. InferenceX and LMSYS data, MTP, DSpark, long context.
Sovereign AI inference: 8 × RTX PRO 6000 Blackwell vs 8 × H200 SXM
Comparative architecture for on-premises AI inference: 8 NVIDIA RTX PRO 6000 Blackwell vs 8 H200 SXM, vLLM throughput, memory bandwidth, and 3-year TCO.
GDPR-compliant AI: the checklist for generative AI in a company
GDPR-compliant AI: what the text requires, what the CNIL and the EDPB recommend, what the EU AI Act adds and when. Sourced checklist, chatbots, RAG and agents included.
Qwen3.8-Flash-Next on a Mac Studio M5 Ultra: 112 GB of weights, one million tokens
180 billion parameters of which 6 active, a measured 111.6 GB in 4-bit MLX, 24 KiB of cache per token: the real sizing on a 512 GB M5 Ultra.
GLM-5.3-Flash on a Mac Studio M5 Ultra: 320 billion parameters within 512 GB
A measured 334.1 GB in 8-bit MLX, 11 KiB of cache per token, a declared one-million-token window: the sizing of GLM-5.3-Flash on an M5 Ultra.
LLM throughput in tokens per second: the four conditions that change everything, measured on sixteen machines
The same model goes from 14 to 290 tokens per second depending on the machine, and software, compression and user count matter as much. Eight sourced tables, explained in plain language.
Mac Studio M5 Ultra: 512GB at 1.2TB/s, what it changes for a local LLM
512GB at 1.2TB/s: what the M5 Ultra changes for a local LLM, which Flash models fit, and what MLX quantisation costs in quality, with measured figures.
AI chatbots and GDPR: the four obligations settled in the architecture
Inform, ground, limit, prove: the four obligations of an enterprise chatbot, and why three are decided in the architecture, not in the terms of service.
Semantic layer and ontology for AI agents
An ontology-based semantic layer gives AI agents governed context. Architecture, knowledge graph, eBPF enforcement and sovereign serverless deployment.
Qwen3.8-Max-Preview: 2.4 trillion parameters, and open questions
Qwen3.8-Max: 2.4 trillion parameters, 95 billion active, and since August 2026 the base model weights published on Hugging Face. What it changes for local LLMs.
Kill switch and local LLMs: what French parliamentary report No. 3054 changes
June 2026: foreign access to two major AI models suspended by US order. Report No. 3054 proposes an open-source tax credit for SMEs.
Orchestrating several coding agents in parallel
Development is moving from the single agent to the fleet: a supervisor delegates to sub-agents, keeps context, and review remains the only real bottleneck.
Installing a local LLM in the enterprise: the guide
Deploying a local LLM in the enterprise: model, NVIDIA hardware certified by HPE, Dell or Supermicro, vLLM or llama.cpp runtime, gateway and security.
Generative AI and GDPR: taking back control
Generative AI and GDPR in 2026: what the CNIL recommends, the extraterritorial risk, and how local AI puts data control back in the hands of the enterprise.
Coding with AI on a local model via LiteLLM
An OpenCode harness, open-weight models served by vLLM, a LiteLLM gateway with quotas: AI coding help without a line of source leaving your infrastructure.
Enterprise chatbot: the internal ChatGPT use case
The enterprise chatbot use case: an internal ChatGPT connected to your documents through RAG, GDPR-compliant and hosted on-premises. Hardware by segment.
What is a sovereign AI platform?
Definition, the difference between on-premises and hybrid, hardware, open-weight models and compliance. The reference guide for scoping a project.
Legal translation and AI: translating a contract without publishing it
Translating a contract online communicates it whole to a third party. Why secrecy is not cured by contract, and how to translate on-premises.
Qwen3.8-27B on-premises: 262,144 tokens of context on a single machine
Gated DeltaNet hybrid architecture, a 64 KiB-per-token KV cache, native 262,144-token context: the real hardware sizing to serve Qwen3.8-27B on site.
Enterprise RAG: connecting AI to your documents
Enterprise RAG in 2026: chunking, hybrid search, reranking and access control. The architecture for connecting an LLM to your documents, on-premises.
Prefill, decode and KV cache: which hardware for an LLM
Prefill is compute-bound, decode is bandwidth-bound, the KV cache grows with every token. Which of the three decides your hardware, from PCIe card to cluster.
Open Knowledge Format: an open format for AI agent memory
OKF represents in Markdown the knowledge AI agents consume. Portable, versionable and free of lock-in, it serves as the memory layer of sovereign platforms.
AI for business applications: CRM, support, ERP, documents
A single internal AI API serves CRM, support, ERP and document management, with virtual keys, budgets and telemetry per application.
The pitfalls of putting an on-premises LLM into production
Eight concrete pitfalls of deploying an on-premises LLM inference service into production, and the measures that prevent each one.