Nox 1.3 Astra (Flagship)
The flagship — deepest reasoning in the family
1.3 Astra is Nox's flagship model. It is a 700-billion-parameter mixture-of-experts reasoning engine that activates 36.9 billion parameters per token, reads a context window of roughly one million tokens, and runs a deliberate reasoning pass before every answer.
Astra sits at the top of the Nox 1.3 generation. It is a reasoning model in the strict sense: before producing an answer, it works through an internal chain of reasoning — laying out factors, testing how they relate, and structuring its conclusion — so its answers reflect deliberate, multi-step work rather than a single fast pass.
The engine is a mixture-of-experts transformer with 700 billion total parameters, of which 36.9 billion are active per token. Routing work to a small set of experts per token is what gives Astra flagship-scale depth without flagship-scale latency. Its context window is roughly 1,000,000 tokens — the largest in the Nox family — so very long conversations and large attached documents stay fully in view without truncation.
Its headline benchmark figure is public: the foundation engine behind Astra has a reported 92.9% on GPQA-Diamond, a benchmark of graduate-level science questions spanning biology, chemistry, and physics, written so that even domain PhDs score far below ceiling. That figure comes from the engine's developer, not from Nox — the numbers Nox itself publishes are its disclosed self-run HealthBench Hard measurement and the measured Leo safety metrics on the accuracy page.
Astra is fully multimodal within Nox: it accepts photo input directly (attached images are handled automatically), can generate charts inline, and can drive Nox's connected-app actions. It also has consent-gated live web access — it asks before searching, and every source it draws on is cited in the answer.
Like every model in the family, Astra runs behind Leo, Nox's deterministic safety layer. Leo screens each message before Astra replies, using a rules-based emergency detector whose recall and false-positive rate are published from a generated evaluation — so the deepest answers in the family pass through exactly the same measured safety gate as the fastest ones.
1.3 Astra is available on MAX. Everyday chat runs the fast models — 1.3 Nova and 1.3 Vela — while Astra is reserved for Nox's deepest work.
Capabilities
- Deliberate multi-step reasoning — an internal chain of thought worked through before every answer
- ~1,000,000-token context window — the largest in the Nox family
- 700-billion-parameter mixture-of-experts engine, 36.9 billion active per token
- Native photo understanding, inline chart generation, and connected-app actions
- Consent-gated live web access with every source cited
Model details
- Architecture — Mixture-of-experts — 700 billion parameters total, 36.9 billion active per token
- Context window — ~1,000,000 tokens
- Reported benchmarks — 92.9% on GPQA-Diamond (graduate-level science) — as reported by the engine's developer
- Reasoning — Deliberate internal reasoning pass worked through before every answer
- Latency — Slower per reply than Nova or Vela — the reasoning pass adds time in exchange for depth
- Photo understanding — Yes — attach photos as usual; Nox handles them automatically
- Live web access — Yes — consent-gated; Nox asks before searching, then cites every source it uses
- Safety layer — Runs behind Leo, Nox's deterministic red-flag detector, on every message
- Available on — MAX
Availability: Available on MAX.