---
license: other
license_name: qwen
license_link: https://huggingface.co/Qwen
base_model:
  - Qwen/Qwen3.5
tags:
  - agent
  - tool-calling
  - function-calling
  - gguf
  - ollama
  - long-context
pipeline_tag: text-generation
library_name: transformers
---

# Engrym Seed Base 9B

Orvyth's seed-tier brain — the recommended member of a local model family built for **tool-using
agents**. Qwen3.5 hybrid linear-attention architecture, **262,144-token native context**, first-class
tool calling.

**Weights are distributed via the Ollama registry.**

```bash
ollama run Orvyth/engrym-seed:base
```

## The ladder

| Tag | Class | Size | 77-task | Tool calls |
|---|---|---:|---:|---:|
| `:nano` | Nano 2B | 2.1 GB | 90.8/143 | 12/12 |
| `:flash` | Flash 4B | 4.6 GB | 124/143 | 12/12 |
| **`:base`** | **Base 9B** | **9.5 GB** | **131/143** | **12/12** |
| `:pro-27b-q4` | Pro 27B v2 Q4 | 16.5 GB | 134/143 | 12/12 |
| `:pro` | Pro 27B v2 Q8 | 28.6 GB | 137/143 | 12/12 |
| `:pro-e` | Pro-E 27B *(experimental)* | 28.6 GB | 137/143 | 12/12 |

## Evaluation

77 tasks · 143 points · `temperature=0` · `max_tokens=16384` · `seed=42` · one attempt · deterministic
validators · **no LLM judge**. Scores are bound to the exact published blobs.

These are **first-party** numbers. Repeated runs on an uncontended GPU are deterministic (zero spread
across n=2 for every model measured), but public reproduction receipts are still pending.

## Compute modes — the score above is a floor

Asking the model to work *deliberately* (reason step by step, verify against every constraint, then
answer) recovers points on tasks it otherwise fails. Base: **131 → 134**. The gain is largest for the
smallest models — Nano gains **+11.8**. On the small end, that is worth more than a model upgrade.

## Defaults

`temperature 0.2` · `top_p 0.9` · `top_k 20` · `num_ctx 32768` · `num_predict 8192`

Native context is 262,144; larger requests are clamped. Default is 32,768 because defaulting to the
native maximum made a 9.5 GB model request ~19 GB of RAM to start. If a prompt exceeds `num_ctx`,
Ollama returns HTTP 400 — it does not silently truncate.

## Lineage

| Stage | Provenance |
|---|---|
| Base | `Qwen/Qwen3.5` — hybrid linear-attention |
| Merge | Ornith-1.0-9B × Qwythos-9B — TIES, 0.5 / 0.5 |
| Tune | Orvyth identity + chip-calling; LoRA merged into the weights |
| Build | Converted and quantized in-house with Orvyth trainkit |

## What is in the artifact

Weights, identity, and tool-call generation. Memory, governed tool execution, safety enforcement,
adapters and multi-agent routing are Orvyth platform concerns, not part of the GGUF. Tool calling is
an output capability — the host validates, authorizes and executes.

## Limits

- Scores are first-party and single-suite. Treat small gaps between adjacent models as unresolved.
- The identity tune is light; under a heavy external system prompt behavior can defer to the base model.
- The 27B is substantially slower per tool call than the 9B. Prefer Base or Flash for agent loops.
- The MTP speculative-decoding head is not included in these builds.
- Tags are mutable — pin the digest for production and evaluations.

---

**ORVYTH** — Intelligence. Governed. *Ground truth over hype. Prove before you claim.*