Models

Each model says what it is for.

What each one does, how much context it holds and what it costs. Then how our training programme works, and where it stands today.

The SynapseX Chat home screen, with the composer, the three modes and the model picker
SynapseX on iPhone: the model picker with Atlas, Forge, Nexus, Oracle, Pulsar and Quasar
SynapseX Chat, screenshot

The models in SynapseX Chat.

  • SynapseX Atlas

    Compact, fast model for general tasks and high volume.

    Try in Chat
    Context
    128K
    Input per 1M tokens
    US$0.13
    Output per 1M tokens
    US$0.65
    • Tools
    • Web search

    Use it for

    General tasks, High volume

  • SynapseX Forge

    Specialized in code generation and review.

    Try in Chat
    Context
    256K
    Input per 1M tokens
    US$0.40
    Output per 1M tokens
    US$1.20
    • Tools
    • Web search

    Use it for

    Code generation, Code review

  • SynapseX Nexus

    Fast multi-agent orchestrator for code, chat, and tools.

    Try in Chat
    Context
    131K
    Input per 1M tokens
    US$0.25
    Output per 1M tokens
    US$0.67
    • Tools
    • Web search

    Use it for

    Multi-agent orchestration, Code, chat and tools

  • SynapseX Oracle

    Advanced reasoning for complex tasks.

    Try in Chat
    Context
    131K
    Input per 1M tokens
    US$2.27
    Output per 1M tokens
    US$4.53
    • Tools
    • Web search
    • Reasoning

    Use it for

    Complex reasoning

  • SynapseX Pulsar

    Multimodal model for documents, visual analysis, charts, and figures.

    Try in Chat
    Context
    200K
    Input per 1M tokens
    US$3.50
    Output per 1M tokens
    US$14.00
    • Tools
    • Web search
    • Vision
    • Reasoning

    Use it for

    Documents, Visual analysis, Charts and figures

  • SynapseX Quasar

    Advanced quantum computing research and frontier reasoning.

    Try in Chat
    Context
    200K
    Input per 1M tokens
    US$3.50
    Output per 1M tokens
    US$14.00
    • Tools
    • Web search

    Use it for

    Quantum computing research, Frontier reasoning

Read from the live SynapseX Chat model list. A price the site could not read shows as unknown, never as a guess.

Use cases

Which one for which work.

Illustration of the product
  1. Atlas is the compact one: general tasks, in high volume, in the language you write.

  2. Forge generates and reviews code. Nexus orchestrates several agents across code, chat and tools.

  3. Pulsar reads documents, charts and figures, and works with what it sees.

  4. Oracle is for complex reasoning. Quasar is for quantum computing research and frontier reasoning.

Recorded in SynapseX

What the answers look like.

  • Human carbonic anhydrase II with acetazolamide, PDB 3HS4, turning in the SynapseX structure viewer

    Turn a protein in 3D

    Human carbonic anhydrase II with acetazolamide, PDB 3HS4, in the structure viewer.

  • A 420 km orbit around the Earth at 51.6 degrees in the SynapseX 3D trajectory viewer

    Follow an orbit

    A 420 km orbit at 51.6 degrees in the trajectory viewer, computed from the closed form.

  • SynapseX Chat answering how the surface code protects a logical qubit, with the threshold formula

    Ask how error correction works

    An answer on how the surface code protects a logical qubit, with its threshold formula.

  • SynapseX Chat comparing the surface code and the color code in a table

    Compare two codes

    Surface code and color code side by side, with a sentence on which to choose.

  • SynapseX Chat writing a typed Python function with doctests

    Write code with its tests

    A typed Python function with doctests and its complexity in one sentence.

How we train

A training programme with a written gate.

Illustration of the product
  1. We do not pre-train from zero. We start from an openly licensed base model, and a model we release names its base and that licence.

  2. Code, science and quantum examples, with licences checked source by source and overlap audited. We do not train on your private research without an explicit, revocable opt-in.

  3. Post-training runs on LUMI-G under EuroHPC allocation EHPC-AIF-2025LS13-019, with GPU-hours budgeted for each stage before it starts.

  4. The same fixed question set goes to the base and to the candidate, and the rule is written before the run. A candidate that loses is recorded as a loss.

  5. A candidate joins the catalogue only if it passes. Until then it stays out, whatever it cost to train.

Where it stands

The programme and the catalogue are separate things. Models from this programme join the catalogue only after they pass the gate. None has so far.

Your data

We do not train our models on your private research without an explicit, revocable opt-in. The full data-use policy is in the privacy policy.