PacketSafari

On-premises deployment · Private AI

Your infrastructure. Your boundary.

Run the same PacketSafari RCA and security workflows inside your controlled infrastructure. Packet storage, identity, retention, AI, and egress follow your deployment policy while conclusions remain inspectable.

Infrastructure, model fit, concurrency, recovery, and workload capacity are confirmed for each deployment profile.

Customer deployment boundaryEgress policy · enforced
Packet storageCapture stays privateRetention · access · audit
PacketSafariCore EngineFrames · flows · fields
Qualified routeYour AI endpointPrivate · hosted · approved API

The complete operating boundary

Control more than the model.

On-premises carries the same PacketSafari evidence contract into a customer-controlled deployment. Qualify the data, identity, AI, and operational controls together.

Packet storage

Full captures and persisted investigation outcomes remain inside the customer-controlled storage boundary.

Identity and access

Connect the deployment to the approved organization identity and access policy.

Retention and audit

Apply customer lifecycle, deletion, evidence-access, and audit requirements.

AI and egress

Route bounded evidence only through an approved customer-operated endpoint and explicit egress policy.

Operations

Qualify infrastructure, updates, recovery, support, and shared operational responsibility before production.

Model + infrastructure sizing

Private-AI deployment planner.

Already own hardware? See which supported models fit. Targeting a model? Compare the reference infrastructure it requires.

Compare capacity fit, PacketSafari support status, context, concurrency headroom, and indicative infrastructure cost. Hardware prices are indicative USD purchase ranges and exclude networking, facilities, support, and PacketSafari. Fast identifies a tested route that met the focused-case target; it is not a universal latency guarantee. Final runtime, topology, and investigation quality are confirmed during deployment qualification.

Supported infrastructure profileModel fit confirmed during deployment

8× H100 · 640 GB

A supported PacketSafari enterprise infrastructure profile for customer-operated AI, with room for larger models and concurrent investigations.

Indicative hardware price$250K–$400K
Supported and evaluated model fitQualification included
  • GPT-OSS-120BDeployment-sized context
    Not recommended

    MXFP4 hardware-fit candidate. Eight GPUs provide capacity for replicas or other workloads, not a measured eightfold speedup.

  • Gemma 4 31BDeployment-sized context
    Not recommended

    Hardware-fit candidate for a shared inference pool. Context, concurrent investigations, and local model quality remain to be qualified.

  • Mistral Medium 3.5FP8 capacity fit context
    Supported · tested

    Fits on a single 4× H100 node at FP8. Additional GPUs can host replicas or increase workload headroom; qualify the exact endpoint and workload.

  • GLM 5.34-bit context
    Supported · Evaluation

    Supported for evaluation; context and concurrency are sized for the deployment.

  • Kimi K24-bit context
    Supported · Fast

    Tested Fast option; usable context and concurrency are sized for the deployment.

  • Kimi K2.6Model-dependent context
    Supported · Fast

    Supported alternative model for a PacketSafari deployment.

  • MiniMax M3MXFP8 capacity fit context
    Not supported

    Capacity fit only; the model failed PacketSafari’s investigation-quality test.

  • DeepSeek V4 Flash 0731Quantized capacity fit context
    Supported · Fast

    Supported focused-model option; exact runtime fit is confirmed during deployment.

  • Qwen3 235BRevision-dependent context
    Not supported

    Capacity fit only.

  • Qwen3 Coder (large)Revision-dependent context
    Not supported

    Capacity fit only.

Frontier deployment comparison

Choose where frontier intelligence runs.

Compare a managed frontier route with supported EU-hosted and private-AI options. The deployment boundary changes; PacketSafari’s packet evidence contract does not.

Managed frontier

OpenAI GPT-6 Sol

Position
Hosted strongest-model route
AI boundary
Approved managed provider
Infrastructure
No customer GPU cluster
Published context
1.05M tokens
Availability
SaaS and approved hosted deployments
Supported EU frontier

Claude Sonnet 5.5 on Amazon Bedrock EU

Position
Frontier quality with EU inference
AI boundary
Customer AWS account, EU inference profiles
Infrastructure
No customer GPU cluster
Published context
200K tokens
Availability
Available now
Supported private reference

Mistral Medium 3.5

Position
Self-hosted on-prem starting point
AI boundary
Customer-operated endpoint, licensed weights
Infrastructure
About one 8× H100 class node
Published context
262K tokens
Availability
After deployment qualification
What the number meansContext is the model’s working envelope, not the PCAP size limit.

PacketSafari keeps packet processing in the Core Engine and supplies compact, relevant evidence to the Agent. The usable context configured for a deployment may be lower than the model’s published maximum.

Shared foundationPacketSafari Core Engine

Bounded packet evidence · inspectable findings · one investigation workflow

Current model evidence

Supported means tested. Not merely listed.

Finding a model through an endpoint is only discovery. PacketSafari separates API compatibility, investigation quality, and the hardware profile that will carry the workload.

Supported · tested

DeepSeek V4 Flash 0731

Self-hostable, EU route available
Great in the model quality index

Open-weights frontier model; reliable Fast investigations in team use.

Supported · tested

DeepSeek V4 Pro 0423

Self-hostable
Great in the model quality index

Open-weights frontier model; reliable investigations in team use.

Supported · tested

Kimi K2.6

Self-hostable
Great in the model quality index

Open-weights frontier model; reliable investigations in team use.

Supported · tested

Kimi K2.7 Code

Self-hostable, EU route available
Great in the model quality index

Open-weights frontier model; reliable investigations in team use.

Supported · tested

GLM 5.3

Self-hostable, EU route available
Good in the model quality index

Open-weights frontier model; completed hard investigation cases with cited evidence.

Supported · tested

Mistral Medium 3.5

Self-hostable, EU route available
Good in the model quality index

Finds the main compromise with cited packets; minor gaps on a full attack chain.

Supported · tested

NVIDIA Nemotron 3.5 Super VL

Self-hostable
Good in the model quality index

Finds the main compromise with cited packets; misses secondary findings.

Not recommended

Gemma 4 31B

Self-hostable
Poor in the model quality index

Runtime parsing failures leave hard cases unanswered.

Not recommended

Llama 3.3 70B Instruct

Self-hostable
Poor in the model quality index

Does not inspect packets; reports only the overview.

Not recommended

Llama 4 Maverick

Self-hostable
Poor in the model quality index

Writes tool calls as text instead of making them; fails qualification.

Not recommended

Llama 4 Scout

Self-hostable
Poor in the model quality index

Writes tool calls as text instead of making them.

Not recommended

gpt-oss-120b

Self-hostable, EU route available
Poor in the model quality index

Fails tool-driven investigation on hard cases.

Hardware-fit references: OpenAI GPT-OSS memory guidance , Google Gemma memory guidance, and Mistral Medium 3.5 model guidance. Loading a model does not establish usable context, concurrency, or PacketSafari investigation quality.

See how self-hostable, non-Chinese models performed on a PacketSafari investigation in the model quality index.

One qualification boundary

Qualify once. Monitor drift without spending.

A short validation uses synthetic data only. Ongoing monitoring checks for material changes without touching a customer capture.

  1. Discover

    Confirm your selected model and deployment profile.

  2. Qualify

    Validate compatibility without using a customer PCAP.

  3. Accept

    Confirm investigation quality and operating capacity.

  4. Monitor

    Keep the approved deployment profile current.

Validated for your deployment. PacketSafari confirms the selected model, endpoint, workload, and capacity profile before production use.

Private AI deployment

Start with your captures, cluster, and security boundary.