ENGRYM SEED

Orvyth's seed-tier brain — a local model family for tool-using agents.

Seed runs local; Epoch and Axiom are hosted tiers.

Not a chatbot. A substrate.

Run Engrym Seed

ollama run Orvyth/engrym-seed

Pulled before 2026-08-09? Re-pull. Ollama does not refresh a tag it already holds, and every tag was rebuilt: a real chat template (older builds shipped a raw passthrough that silently discarded system prompts and made tool calling unavailable), a sane default context, and a header repair that fixed artifacts which would not load at all.

The Seed ladder

Experimental channel

-e tags are the staging channel: internal evaluation, never :latest, published so results can be reproduced rather than because they are finished. "E" builds are trained on multi-step tool loops — their advantage is step efficiency (reaching a goal in fewer tool calls), not raw knowledge.

📊 Measured evaluations

These are first-party numbers. One ruleset applied identically to every row: 77 tasks · 143 points · temperature=0 · max_tokens=16384 · seed=42 · one attempt per task · deterministic validators · no LLM judge. Runner: Ollama on a single NVIDIA B200. Scores are bound to the exact published blobs, not to an internal build.

Read these as single-run figures, not a ranking. Each row is one run. We have measured meaningful run-to-run variation on this suite — a task timeout that fires under load can cost a large model double-digit points — so differences of a few points between adjacent rows are not evidence that one model is better than another. A repeated-run study is in progress and these numbers will be restated with spread when it completes. Public reproduction receipts are pending.

Compute modes are real, and the numbers above are floors

Every score above was produced with one terse system prompt at the model's default effort. Asking a model to work deliberately — reason step by step, verify against every constraint, then answer — recovers points on tasks it otherwise fails. Measured by retrying only the failed tasks under a deliberate-mode prompt:

The gain is largest for the smallest models. Pro-E gains nothing because it already deliberates by default; Nano gains twelve points because it answers quickly unless told otherwise. If you want accuracy over latency, say so in your system prompt — it is worth more than a model upgrade on the small end.

Tool calling — measured under a realistic agent load

Tested through Ollama's /api/chat tools interface with a 45-tool schema and a full agent system prompt, across three scenarios, 12 trials each. A trial passes only if the response populates tool_calls; prose describing the action counts as a failure. Every published build scored 12/12, in both a 2-tool short-prompt profile and the 45-tool profile.

Context and memory — measured

Native context is 262,144 tokens. Larger requests are silently clamped: ask for 524,288 and Ollama logs requested context size too large for model, n_ctx_train=262144 and gives you 262,144. There is no 512K mode — that would require RoPE extension and retraining.

Context is cheap here because most layers are linear-attention: 32× the window for roughly 2× the memory. The default is 32,768 rather than the native maximum because defaulting to 262,144 made a 9.5 GB model request roughly 19 GB of RAM just to start, and the 27B would not load at all on a 32 GB machine.

If your prompt exceeds num_ctx, Ollama returns HTTP 400 (request (N tokens) exceeds the available context size) — it does not silently drop part of your prompt. A 45-tool schema plus an agent system prompt is roughly 6–8k tokens, so it fits comfortably at the shipped default. Raise num_ctx for long conversation histories, not for tool schemas.

Sampling defaults

Defaults are per class, chosen from measured temperature sensitivity rather than a single house value.

The E builds run cooler because the tools-trained line is measurably less temperature-robust: it loses roughly 13 points between temperature 0 and 1.0, where the plain Base line holds within a few points. Benchmarks are run at temperature=0.

What changed in this release

Corrections

Claims previously published here that did not hold up, withdrawn rather than quietly edited:

• The V1.1a "86.7%, tied #2" composite and its 58-model leaderboard. Produced by an artifact that does not load, graded against two incorrect gold answers. Void.
• GSM8K / IFEval / HellaSwag / WinoGrande / ARC figures. Measured on artifacts and precisions that are no longer what ships here.
• "VRAM @ 262K = 14 GB." Measured at 18 GB.
• "The 9B is a reasoning model, not a tool driver." False — it was true only while the artifact shipped no chat template.
• "A small num_ctx truncates the tool schema out of the prompt." Incorrect. Ollama returns HTTP 400 for an oversized prompt; it does not truncate. Tool calling was verified working at 32768, 8192 and 7000.
• A published table attributing a 1/12 tool-call result to num_ctx 8192. That figure came from a third party whose harness explicitly requested 65536. It was never measured at 8192, and the cause of that result remains unexplained.
• :v1 described as a 5.6 GB Q4 artifact scoring 85.3%. The published :v1 is an 18 GB f16 build that has never been run against the corrected suite. No score is claimed for it.

Model profile — Base 9B

Lineage

OpenAI-compatible API

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

What is in the artifact, and what is not

This is the bare seed brain: weights, identity, and tool-call generation. Memory, governed tool execution, safety enforcement, adapters, voices and multi-agent routing are Orvyth platform concerns and are not in the GGUF. Tool calling is an output capability — the host remains responsible for tool definitions, argument validation, authorization, execution and result delivery.

The Engrym line

ORVYTH

Intelligence. Governed.

Ground truth over hype. Prove before you claim.