engrym-seed-flash-4b

Engrym Seed Flash 4B

Part of Engrym Seed — Orvyth's seed-tier local model family for tool-using agents. Qwen3.5 hybrid linear-attention architecture, 262,144-token native context, first-class tool calling.

The best agent build — fastest to act, and the value pick of the ladder.

Weights are distributed via the Ollama registry.

ollama run Orvyth/engrym-seed:flash

The full ladder

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.

First-party numbers. Repeated runs on an uncontended GPU are deterministic (zero spread across n=2 for every model measured). Receipts: https://huggingface.co/datasets/Orvyth/engrym-seed-receipts

Compute modes — the score is a floor

Asking the model to work deliberately recovers points on tasks it otherwise fails, and the gain is largest for the smallest models (Nano +11.8, Flash +6, Base +3, Pro-E +0). 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

If a prompt exceeds num_ctx, Ollama returns HTTP 400 — it does not silently truncate.

Lineage

Qwen/Qwen3.5 base → Ornith-1.0-9B x Qwythos-9B TIES merge (9B line) → Orvyth identity and chip-calling LoRA merged into the weights → converted and quantized in-house.

Limits

• First-party scores; treat small gaps between adjacent models as unresolved.
• The identity tune is light; a heavy external system prompt can pull behavior toward the base model.
• The 27B is much slower per tool call than the 9B and smaller.
• No MTP speculative-decoding head in these builds.
• Tags are mutable — pin the digest for production and evaluations.

---

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