ENGRYM SEED

Local open-weight model artifacts for Ollama

Engrym Seed is Orvyth’s published local model family. Choose an explicit tag, run it on your own machine, and inspect the fixed public record behind a cited artifact or evaluation result.

Seed is available as local Ollama artifacts. There is no public Orvyth hosted runtime, chat service, account service, billing program, or generic API to sign up for today.

Start with the record, not a mutable alias. For a normal local run, use Orvyth/engrym-seed:base rather than an unqualified family name or :latest.

Project, artifact, and evidence records

Quick start: Base 9B

ollama pull Orvyth/engrym-seed:base
ollama run Orvyth/engrym-seed:base

base is the explicit 9B artifact recorded in this document. At the receipt revision below, :base, :base-v1.2ag1, and :latest point to the same model-layer digest. Tags remain mutable, so inspect the tag you actually pull.

Receipt snapshot and reproducibility boundary

This card cites the following fixed Hugging Face revision:

90406a570e988e8d2304cdf2d9c367678119c4da

The revision contains the recorded model-layer digests, task suite, runner configuration, score summaries, and raw per-task artifacts. Each score below is scoped to the named tag and run inside that revision.

Ollama tag names can change after this card is published. A matching tag name is not proof that a later pull resolves to the same blob. For reproducible work:

1. Pull or inspect the exact named tag.
2. Compare its model-layer digest with published_digests.json.
3. Keep the resolved artifact identity with your own run record.

Featured current artifacts

The following table is the primary local ladder. Scores are one first-party current-blob run under the shared evaluation rules documented later in this card. The one-run latency column is descriptive of that recorded environment; it is not a general throughput or hardware-performance claim.

Q4 and Q8 are different artifacts

pro-27b-q4 and pro are not interchangeable names for the same 27B file. They have different quantization, byte size, and model-layer digest, even though their separately recorded one-run scores happen to match.

Experimental artifacts

Experimental tags are published for inspection and reproducibility. They are not the default local-start recommendation, and their score entries should be read with the same first-party, single-run limits as the main table.

pro-e-v2 and pro-e-v2-text are public experimental artifacts with no score claim in this card. Do not infer an evaluation result from a tag’s presence.

Recorded model-layer digests

These are the full model-layer digests recorded at the receipt revision cited by this card. They are included here for auditability; the registry’s live tag page remains the source to inspect before a new pull.

Tag map, aliases, and historical artifacts

The registry currently contains multiple names for some recorded blobs. The map below records their intent at the cited receipt snapshot, not an immutable future registry state.

The historical :v1 artifact retains a different, raw prompt template. It is not evidence that every public Engrym tag shares the current tool-aware template or capability profile.

Evaluation protocol

All score rows above cite the same first-party current-blob protocol at the pinned receipt revision.

The configuration is public in config/settings.json.

What the scores establish

They establish that the named receipt revision records the named model-layer digest and a named one-run result under the protocol above.

What the scores do not establish

• An upstream-base delta or merge benefit, because no direct upstream baseline is published in this record.
• Independent reproduction, because the cited runs are first-party.
• A cross-product ranking, leaderboard position, or general intelligence rank.
• Agent-loop reliability, tool-execution reliability, safety certification, or application-level authorization behavior.
• A general latency, throughput, VRAM, or hardware recommendation for another machine or deployment configuration.

Base 9B profile

This section applies only to Orvyth/engrym-seed:base and its recorded alias group. It is not a family-wide provenance or license statement.

The evaluation protocol uses temperature 0.0; that is different from the current tag’s normal sampling defaults.

Context, chat, and tool boundary

The current featured tags advertise a native 256K context window. The shipped Base 9B default is num_ctx=32768. Treat context capacity and local memory use as tag- and host-specific operational properties. Inspect the exact artifact and test it in the environment you plan to use.

Featured current tags use the current Ollama chat template recorded in the receipt snapshot. The host application remains responsible for every real tool operation: tool definitions, argument validation, authorization, execution, result delivery, logging, and user-facing safety controls.

Example local OpenAI-compatible call:

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Orvyth/engrym-seed:base",
    "messages": [
      {"role": "user", "content": "Hello"}
    ]
  }'

Base 9B lineage and license scope

The Base 9B line declares the following provenance:

The Base 9B artifact is a Qwen3.5 derivative and follows the applicable upstream base-model licensing terms. Inspect the license and metadata attached to the exact tag before use. This statement does not apply automatically to Nano, Flash, Pro, experimental, or historical tags.

Artifact boundary

Engrym Seed is a local model artifact. It does not itself provide:

• Hosted inference or a public Orvyth API
• Persistent memory or a user account system
• Governed tool execution or application authorization
• Safety enforcement beyond what a host application implements
• Voice, multi-agent routing, billing, or a SaaS product

Release and correction policy

This card is intentionally a release record, not a leaderboard. When a tag, template, digest, benchmark, or claim changes, update the named source record and retain a dated correction rather than silently carrying old claims forward.

Historical artifacts remain available for compatibility and inspection. A historical or experimental tag without a named current-blob receipt has no performance claim on this card.

Citation-ready references

• Ollama family page
• Base 9B tag
• Published digests at the fixed receipt revision
• Result index at the fixed receipt revision
• Orvyth release manifest

Intelligence. Governed. Ground truth over hype. Prove before you claim.