Extremely detail-oriented
Specs, edge cases, service contracts, tests, and fine copy. Luna is not happy with “it works” — it leaves work ready for a professional review.
Luna series · The detail specialist FRONTIER
Xeretron Luna 304B
The model you will want on your machine. Not for vibe coding: for architecture, key decisions, and consulting its wisdom. A large model in this class is not used to write code — it is consulted. Runs at 12–18 t/s depending on the GPU. On the Artificial Analysis Intelligence Index it competes with GPT-5.6 Luna Max, outperforms DeepSeek V4 Pro, sits near GLM-5.3, and beats Kimi K2.6 on intelligence — on your node, with no rented tokens.
Local throughput: 12–18 t/s depending on GPU, quantization, and context (Xeretron measurement). AA positioning vs the public models cited. Luna is not the model that “writes the PR” — it is the one you consult for architecture and hard decisions. Verify benches on artificialanalysis.ai.
Luna strengths
Detail that hurts… because it is that good.
Not for vibe coding. The model you consult when the architecture call matters — and Dev writes the code.
Professional at architecture
Clear plans, domain boundaries, explicit trade-offs, and decisions that hold in production. When the system matters more than the snippet.
Competes with GPT-5.6 Luna Max
On the Artificial Analysis Intelligence Index, GPT-5.6 Luna Max is the cloud “Luna” reference. Xeretron Luna 304B is built to fight in that league — on your machine.
Outperforms DeepSeek V4 Pro
DeepSeek V4 Pro 0813 (Max) scores 36 on the Intelligence Index. Luna is positioned above that open bar for detail and architecture work.
Close to GLM-5.3
GLM-5.3 (Max) leads open weights at 45 on the Index. Luna sits in that neighborhood of intelligence — with the control and privacy of your Xeretron node.
More intelligent than Kimi K2.6
Kimi K2.6 (reasoning) scores 27 on the Index (~24 non-reasoning). Luna aims clearly above: more depth for architecture and demanding decisions.
Principal Architect in CodiiX
In CodiiX, Luna is your Principal Architect and advisor to the Dev model on very complex cases: plans, trade-offs, and criteria — Dev writes; Luna counsels with wisdom.
No per-token cost
Runs on your GPU at 12–18 t/s. Consult architecture and hard decisions without watching a meter. The bill is your hardware — not a Frontier SaaS.
Built for your node
OpenAI-compatible API, profiles in the Xeretron panel, VRAM/context evaluation before promising miracles. Usable Frontier — not just “loadable”.
Omega 320B · brain of brains
Frontier orchestration
- 320B · GLM-5.3-Flash base
- Agents, devices, impossible bugs
- ~18 t/s on 2× RTX 3090
- Ideal for fleets and systems
Luna 304B · the detail specialist
When architecture is the mission
- 304B · Frontier class
- 12–18 t/s depending on GPU
- Architecture & key decisions — not vibe coding
- Principal Architect / Dev advisor in CodiiX
Why Luna
The moon of your stack: clarity under pressure.
A large model in this class is not used to write code — it is used to consult its wisdom. Luna is precise on architecture; daily craft stays with the Dev model.
Architecture and key decisions
Domain boundaries, trade-offs, risks, and design criteria. When the question is “what system do we build?”, not “what function do I write?”.
Advisor to Dev
In CodiiX, Luna acts as Principal Architect: advising the Dev model on very complex cases. Wisdom above; implementation below.
Wisdom, not vibes
Not the “ship code at 70 t/s” model. The one you consult when the problem is deep: multi-service, multi-constraint, multi-stakeholder.
Privacy by design
Your plans and specs do not leave to someone else’s Frontier API. Luna lives inside your perimeter.
Artificial Analysis · Intelligence Index
The map where Luna competes.
Public Artificial Analysis Intelligence Index (v4.3.2) scores for the models Luna is compared against. Source: artificialanalysis.ai — including DeepSeek V4 and OpenAI’s Luna line.
| Public model | Intelligence Index | Notes |
|---|---|---|
| GLM-5.3 (Max) | 45 | Top open-weights reference · Luna sits nearby |
| GPT-5.6 Luna (Max) | ~37 | Proprietary “Luna” reference · Luna 304B competes here |
| DeepSeek V4 Pro 0813 (Max) | 36 | Luna positioned above (outperforms) |
| Kimi K2.6 (reasoning) | 27 | Luna surpasses on intelligence (Index; ~24 non-reasoning) |
| Xeretron Luna 304B | Your node | Architecture + wisdom · 12–18 t/s · no rented tokens |
*Estimated / non-reasoning on AA. The Index combines 10 evaluations (AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, HLE, GDP.pdf, CritPt, AA-Omniscience, AA-LCR). Scores move with each refresh. See also DeepSeek V4 Flash and GLM-5.3 vs DeepSeek V4 Pro.
Specifications
Luna technical sheet.
| Attribute | Detail |
|---|---|
| Name | Xeretron Luna 304B |
| Class | ~304B · local Frontier · detail flagship |
| Positioning | Competes with GPT-5.6 Luna Max · > DeepSeek V4 Pro · ≈ GLM-5.3 · > Kimi K2.6 (AA Intelligence Index) |
| Specialty | Architecture · key decisions · consult wisdom (not vibe coding / not code writing) |
| Ideal use | Principal Architect in CodiiX · advisor to the Dev model on very complex cases · system design |
| Local throughput | 12–18 t/s depending on GPU (Xeretron measurement; varies by quantization/context) |
| Deployment | Xeretron node · OpenAI-compatible API · prior hardware evaluation |
| Privacy | Local / on-premise · no third-party token costs |
| Bench reference | Artificial Analysis |
License & access
Luna ships with your Xeretron plan / evaluation. We confirm quantization profile, VRAM, and usage terms during onboarding.
On your node
Definitely a model you will want on your team.
Luna 304B runs at 12–18 t/s depending on the GPU. We size quantization, context, and lanes with you — so the Frontier Architect is actually consultable, not just “loadable”.
- Luna profile · Principal Architect in CodiiX.
- Advisor to the Dev model on very complex cases.
- 12–18 t/s depending on GPU · VRAM/context evaluation.
- No rented tokens: consult without fear.
Will it fit my hardware?
Tell us GPUs, RAM, and use cases. We will say with numbers whether Luna is the sweet spot — or if Turbo / Titan / Omega fits better.
Next step
Bring the moon to your node.
Luna for architecture and key decisions. Developer to write code. Omega to orchestrate. Titan for product perfection.