Nox Orion 2.1 (Flagship)

Frontier AI health model · Released in Nox: October 22, 2026 · Page last updated: October 24, 2026

The flagship, now with a peer-review answer pipeline

Orion 2.1 is the flagship of the Nox 2 generation and the successor to Orion 2. The engine is unchanged — an 824-billion-parameter mixture-of-experts model, 46 billion active per token, with a roughly one-million-token context window — and the change is in how the answer is produced: a peer-review pipeline that briefs the case, writes a specialist checklist before drafting, drafts against it, and has the draft reviewed before anything reaches you. It was designed from the point losses measured on HealthBench Professional: missing named specifics, embedded errors in pasted plans that went unflagged, decisive context that was never asked for, mixed-language replies, and over-long answers.

Architecture: mixture-of-experts, 824 billion parameters total, 46 billion active per token — the 700-billion-parameter backbone of 1.3 Astra extended with roughly 100 billion parameters of additional trainable components. Orion 2.1 runs the same engine as Orion 2.

Peer-review pipeline (new in 2.1): before drafting, Orion 2.1 writes a task brief — deliverable, audience, language, constraints, claims embedded in the user's own text, missing context that would change the advice, and the named specifics a specialist would expect — and, in parallel, a specialist checklist. The draft is written against both, then an independent review pass checks the draft item by item and returns the corrected reply. Only the final reply is shown; the review is never pasted into the answer.

Length: earlier Nox models answered under a fixed 4,000-character cap. Lifting that cap alone raised HealthBench Hard by 8.4 ± 2.0 points on a 150-example paired test, almost entirely on completeness. Orion scales length to the question instead — short for simple questions, up to 9,000 characters for multi-part ones — and covers the direct answer, main alternatives, next steps, and what would warrant urgent care.

Measured on the 1.4 Orion pipeline in September 2026: HealthBench Hard, Nox self-run, official open protocol, all 1,000 examples, scored 46.4% (±0.8) at Ultra and 53.3% (±1.0) at Hyper Ultra. Orion 2.1 runs the same engine, so these remain the most recent published Hard results; they have not been re-measured under the Orion 2.1 name. 1.3 Astra on the identical protocol: 34.8%. Separately, the evaluated Orion 2.1 + compact Peer Review configuration completed all 525 HealthBench Professional tasks on October 23, 2026: raw 72.8%, length-adjusted 69.5%, using GPT-5.4 low as grader. This Nox self-run includes live web evidence and failure-only recovery of three tasks; it is not an independent certification.

Hyper Ultra is Orion’s sixth and deepest thinking level, in Plan mode on MAX. Same model, several passes over one question — draft, self-written rubric, self-grade, revision — about five to six minutes per answer. The 1.4 Orion pipeline scored Light 36.2%, Balanced 36.8%, Extra 36.9%, and Max 38.5% on a fixed 150-prompt development subset.

Multimodal and tool-using: up to ten photos per message; live web access in Plan mode with up to ten search-and-read rounds per answer, every source cited; inline health and anatomy image generation; thinking time shown on every answer and a live step timeline in Plan mode. In NOX WORK it also reads attached documents (PDF, Word, Excel, PowerPoint, CSV, text, code — three files of 15 MB each), removes backgrounds from and upscales images, and reads connected apps.

Runs behind Leo, Nox’s deterministic safety layer, on every message; web, image, and app actions are withheld on any turn Leo flags as a possible emergency.

Available in Plan mode on Pro and MAX (Hyper Ultra on MAX) and as a selectable model in NOX WORK. Not an everyday-chat model.

Capabilities

824-billion-parameter mixture-of-experts model, 46 billion active per token
~1,000,000-token context window
Deliberate internal reasoning pass before every answer, thinking time shown
Six thinking levels, Light to Hyper Ultra — Hyper Ultra (Plan mode, MAX): draft, self-written rubric, self-grade, revision
Dynamic answer length, up to a 9,000-character cap
Photo input — up to ten images per message
Live web access in Plan mode — up to ten search-and-read rounds per answer, every source cited
Image generation — health and anatomy illustrations inline
Plan mode progress timeline
In NOX WORK: document reading, image background removal and upscaling, connected-app reads

Model details

Architecture — Mixture-of-experts — 824 billion parameters total, 46 billion active per token; the 700-billion-parameter 1.3 Astra backbone extended with roughly 100 billion parameters of additional trainable components.
Context window — ~1,000,000 tokens
Answer length — Dynamic — sized to the question, up to a 9,000-character hard cap
Reasoning — Deliberate internal reasoning pass before every answer, with thinking time shown
Answer pipeline — Peer review (new in 2.1) — task brief and specialist checklist before drafting, draft against both, independent review pass; only the final reply is shown
HealthBench Hard — Measured on the 1.4 Orion pipeline (same engine): 46.4% (±0.8) at Ultra, September 5, 2026; 53.3% (±1.0) at Hyper Ultra, September 11, 2026 — Nox's own self-runs, official open protocol, all 1,000 examples; not re-measured under 2.1; not independently verified.
HealthBench Professional — Orion 2.1 + compact Peer Review: Nox self-evaluation, October 23, 2026, all 525 tasks completed. Raw 72.8%; length-adjusted 69.5%; 485 tasks held out of tuning scored 69.8% adjusted. GPT-5.4 low grader; single run with failure-only recovery of three tasks, live web evidence and multi-stage review. Forty tasks were used during development. This measures the evaluated configuration, not every production request; not independently verified or a clinical certification. Summary and methodology: /data/healthbench/orion-2-peer-review/README.md
Thinking levels — Six — Light, Balanced, Extra, Max, Ultra, Hyper Ultra (Hyper Ultra in Plan mode on MAX, ~5–6 minutes per answer)
Raw evidence — Every answer, every grader verdict, aggregate files and SHA-256 checksums are published: Ultra at https://www.getnox.ai/data/healthbench/orion/README.md, Hyper Ultra at https://www.getnox.ai/data/healthbench/orion-hyper-ultra/README.md
Photo understanding — Yes — up to ten images per message, read automatically
Live web access — Yes — in Plan mode, up to ten search-and-read rounds per answer, every source cited
Image generation — Yes — health and anatomy illustrations rendered inline on request
Safety layer — Runs behind Leo, Nox's deterministic red-flag detector, on every message
Available on — Plan mode on Pro and MAX; selectable in NOX WORK

Availability: Available in Plan mode on Pro and MAX.

Nox Orion 2.1 is the named replacement for Nox Orion 2 (retired on October 22, 2026), Nox 1.3 Astra (retired on October 5, 2026), Nox 1.4 Orion (retired on September 23, 2026), Nox 1.3 Nova (retired on September 23, 2026), Nox Agent 1.2 (retired on September 23, 2026).

Nox Orion 2.1 — frequently asked questions

What is Nox Orion 2.1?

Nox Orion 2.1 is Nox's frontier ai health model. Orion 2.1 is the flagship of the Nox 2 generation and the successor to Orion 2. The engine is unchanged — an 824-billion-parameter mixture-of-experts model, 46 billion active per token, with a roughly one-million-token context window — and the change is in how the answer is produced: a peer-review pipeline that briefs the case, writes a specialist checklist before drafting, drafts against it, and has the draft reviewed before anything reaches you. It was designed from the point losses measured on HealthBench Professional: missing named specifics, embedded errors in pasted plans that went unflagged, decisive context that was never asked for, mixed-language replies, and over-long answers.

What can Nox Orion 2.1 do?

Nox Orion 2.1's capabilities: 824-billion-parameter mixture-of-experts model, 46 billion active per token; ~1,000,000-token context window; Deliberate internal reasoning pass before every answer, thinking time shown; Six thinking levels, Light to Hyper Ultra — Hyper Ultra (Plan mode, MAX): draft, self-written rubric, self-grade, revision; Dynamic answer length, up to a 9,000-character cap; Photo input — up to ten images per message; Live web access in Plan mode — up to ten search-and-read rounds per answer, every source cited; Image generation — health and anatomy illustrations inline; Plan mode progress timeline; In NOX WORK: document reading, image background removal and upscaling, connected-app reads.

What architecture does Nox Orion 2.1 use?

Architecture: Mixture-of-experts — 824 billion parameters total, 46 billion active per token; the 700-billion-parameter 1.3 Astra backbone extended with roughly 100 billion parameters of additional trainable components.

How large is Nox Orion 2.1's context window?

Context window: ~1,000,000 tokens

How long are Nox Orion 2.1's answers?

Answer length: Dynamic — sized to the question, up to a 9,000-character hard cap

Which thinking levels does Nox Orion 2.1 offer?

Thinking levels: Six — Light, Balanced, Extra, Max, Ultra, Hyper Ultra (Hyper Ultra in Plan mode on MAX, ~5–6 minutes per answer)

Which Nox plans include Nox Orion 2.1?

Available in Plan mode on Pro and MAX.

Which retired Nox models does Nox Orion 2.1 replace?

Related Nox models

Nox Orion 2 — Retired — the first flagship of the Nox 2 generation (retired)
Nox 1.3 Astra — Deep reasoning with a one-million-token context — retired October 5, 2026 (retired)
Nox 1.4 Orion — Retired — the flagship of the Nox 1.4 generation (retired)
Nox 1.3 Nova — Retired — the balanced model of the 1.3 family (retired)
Nox Agent 1.2 — Retired — Nox's former dedicated research mode (retired)
Nox Orion Flash 2.1 — The fast Orion on a new engine — 552B mixture-of-experts, native photo input
Nox Vela 2.1 — The fast model on every plan
Nox Voice 2 — The most powerful Nox voice yet — expressive, interruptible, Plan mode in the call

Back to the Nox model family · Plans and pricing · Nox features