CheckCom AI guide

Understanding AI models: new LLMs, quantization and local AI

New LLM versions, API routes, open-weight models, quantized releases and local hardware options appear continuously. CheckCom explains what matters for enterprises and how price, source quality, local operation and procurement risk fit together.

Why this guide matters

Almost every day, new AI models, preview versions, multimodal systems, coding models or open-weight variants appear. For enterprises, not every new model is relevant. What matters is availability, cost, source transparency, operational fit and suitability for the intended use case.

public release candidates

condensed from news, model catalog and release signals.

850
local/quantized models

shown as candidates with measurement status, hardware and source notes.

1477
price rows

with EUR normalization, source quality and review status.

Contents

New modelsQuantizationLocal operationHardware profilesUpcoming measurementsMore CheckCom pages

New LLM versions: what enterprises should check

New model versions may bring better quality, lower costs, larger context windows, multimodal capabilities, coding features or new API access. At the same time, they may be previews, regionally restricted, priced differently or replacements for older models.

Key questions

  • Is it stable or a preview?
  • Is there an official provider source?
  • Is API access, open-weight download or an aggregator route available?
  • Are public prices available?
  • Is there specific EU/DACH procurement relevance?

CheckCom link

The model release radar connects AI news, the model catalog, pricing data and release signals. This helps identify which announcements are model-, price- or availability-relevant.

Open model release radar

Quantized models explained

Quantization means that an AI model is stored and executed in a more memory-efficient format. This can make larger models usable on local hardware such as workstations, mini PCs or specific edge setups. Common terms include GGUF, AWQ, GPTQ, EXL2 and quantization levels such as Q4, Q5, Q6 or Q8.

TermSimple explanationEnterprise note
Q4heavily compressed, often memory-efficient and fastgood entry point; quality must be checked for the use case
Q5 / Q6often a strong compromise between size and qualityinteresting when enough VRAM/RAM is available
Q8closer to higher precision, but largeruseful for smaller/mid-sized models or larger hardware profiles
GGUFcommon in the llama.cpp/Ollama/LM Studio ecosystemuseful for inventory and future benchmarks
AWQ/GPTQ/EXL2runtime-specific quantization formatscomparability depends strongly on runtime and settings

A quantized model is not automatically better or cheaper. Memory needs, speed, response quality, context length, license, maintenance and operating costs must be considered together.

Running AI locally: advantages and disadvantages

Advantages

  • more control over data paths
  • less dependency on external API providers
  • potentially more predictable costs at high utilization
  • offline and edge scenarios are possible
  • experimentation without ongoing API costs

Disadvantages

  • hardware, power and maintenance costs
  • model management and updates
  • limited concurrency
  • technical operating responsibility
  • quality and speed vary significantly by model

Local AI is not automatically cheaper or better. It is most relevant when data control, predictable usage, edge automation or local experimentation are important.

Raspberry Pi, workstation, Apple Studio or API?

Different hardware profiles serve different needs. A Raspberry Pi 5 with AI HAT+ 2 is interesting for edge AI and small local tasks, but it is not a large GPU workstation. An RTX 5090 workstation is better suited for larger quantized models and local RAG/coding tests. API and cloud offerings remain important for scaling, enterprise contracts and high concurrency.

ProfileSuitable forLimits
Raspberry Pi 5 CPU-onlysmall tests, light local models, edge demosslow for larger LLMs
Raspberry Pi 5 + AI HAT+ 2edge AI, small LLM/VLM tasks, vision, demo scenariosnot a large LLM workstation
RTX 5090 32 GBlarger quantized models, coding, local RAG testslimited for very large models and long contexts
RTX Pro 6000 / 96 GB classlarger models, long contexts, professional local inferencehigh hardware and operating costs
Apple Studio / unified memorylocal models with large shared memorydifferent runtime and performance profiles
API / Cloudscaling, enterprise, low initial investmentongoing costs, data path, vendor dependency

Upcoming CheckCom measurements

CheckCom is preparing additional measurement series for local AI. Planned profiles include RTX 5090, professional GPUs such as RTX Pro 6000, Apple Studio and Raspberry Pi 5 with AI HAT+ 2. Future results will clearly mark whether a value is externally researched, derived from metadata or measured by CheckCom.

Until CheckCom measurements are available, local model data is marked as metadata, external benchmark reference or not yet measured by CheckCom. No guaranteed token rates are published without measurement status.