ENGRYM SEED

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

Not a chatbot. A substrate.

sizes 2B · 4B · 9B · 27B context 262,144 native quant Q8_0 tool calling 12/12 verified arch Qwen3.5 hybrid

Get it on Ollama All Orvyth models

Run it

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.

Older builds shipped {{ .Prompt }} as their chat template — a raw passthrough that silently discarded system prompts and made tool calling unavailable. Several artifacts also declared one more transformer block than they contained and would not load at all. Both are fixed.

The ladder

Tag	Class	Size	77-task	Tool calls
`:nano`	Seed Nano 2B	2.1 GB	90.8/143	12/12
`:flash`	Seed Flash 4B	4.6 GB	124/143	12/12
`:base` · `:latest`	Seed Base 9B	9.5 GB	131/143	12/12
`:pro-27b-q4`	Seed Pro 27B v2 (Q4_K_M)	16.5 GB	134/143	12/12
`:pro`	Seed Pro 27B v2 (Q8_0)	28.6 GB	134/143	12/12
`:pro-e`	Pro-E 27B (experimental)	28.6 GB	137/143	12/12
`:v1`	Legacy V1 9B (f16)	18 GB	not measured	—

Start with :base for the best capability-per-gigabyte, or :flash if you want the fastest agent loop — it reaches 124/143 at 4.6 GB.

Experimental channel

-e tags are staging: 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. :pro-e-v2, :pro-e-v2-text and :pro-e-9b are also published; :pro-e-9b is an archived early attempt and scores well below the stable line.

Measured evaluations

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. Scores are bound to the exact published blobs.

These are first-party, single-run numbers — not a ranking. We have measured meaningful run-to-run variation on this suite: a task timeout firing under load can cost a large model double-digit points. 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 will be restated with spread. Public reproduction receipts are pending.

Artifact	Score	Percent
Pro-E 27B `:pro-e`	137.0 / 143	95.8%
Pro 27B v2 `:pro`	134.0 / 143	93.7%
Base 9B `:base`	131.0 / 143	91.6%
Flash 4B `:flash`	124.0 / 143	86.7%
Pro-E 9B `:pro-e-9b`	105.0 / 143	73.4%
Nano 2B `:nano`	90.8 / 143	63.5%

Compute modes are real — and the scores above are floors

Every score above used 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:

Artifact	Default	Deliberate	Gain
`:pro-e`	137	137	+0
`:pro`	134	137	+3
`:base`	131	134	+3
`:flash`	124	130	+6
`:pro-e-9b`	105	117	+12
`:nano`	90.8	102.6	+11.8

The gain is largest for the smallest models. Pro-E gains nothing — it already deliberates by default. Nano gains twelve points because it answers quickly unless told otherwise. On the small end, asking for deliberation in your system prompt is worth more than a model upgrade.

Tool calling

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

Native context is 262,144 tokens, and it is enforced: larger requests are silently clamped (n_ctx_train=262144). There is no 512K mode — that would need RoPE extension and retraining.

`num_ctx`	Base 9B footprint
8,192	9.0 GB
16,384	9.2 GB
32,768 — shipped default	9.8 GB
65,536	11 GB
131,072	13 GB
262,144 — native max	18 GB

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 about 19 GB of RAM just to start.

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

Sampling defaults

Setting	Nano · Flash · Base · Pro	E builds
`temperature`	`0.2`	`0.1`
`top_p`	`0.9`	`0.9`
`top_k`	`20`	`20`
`num_ctx`	`32768`	`32768`
`num_predict`	`8192`	`8192`

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 stable line holds within a few points.

Lineage

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

Corrections

Claims previously published on Orvyth surfaces 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.
"VRAM @ 262K = 14 GB." Measured at 18 GB.
"The 9B is a reasoning model, not a tool driver." False — 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 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 its cause remains unexplained.
:v1 described as a 5.6 GB Q4 artifact scoring 85.3%. The published :v1 is an 18 GB f16 build never run against the corrected suite. No score is claimed for it.
Two grader defects in our own benchmark. One made 9 of 143 points unwinnable for every model ever run; another graded a model's explanation instead of its answer. Both fixed, with regression tests. All affected scores were re-graded.

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.

Honest limits

Constraint	Ground truth
Evidence	All scores are first-party and single-run. Public reproduction receipts are pending. Treat small gaps between adjacent models as unresolved.
Tier	Seed-tier is the fast, cheap, always-on layer. Route hard problems up-tier.
Identity	The identity tune is light by design. Under a heavy external system prompt, behavior can defer toward the base model.
Reasoning budget	These are reasoning models. A harness that caps output too low truncates mid-thought and returns prose with no tool call.
27B latency	Substantially slower per tool call than the 9B and smaller. Prefer Base or Flash for agent loops.
E builds	Tool-loop specialists. Step-efficient, but less temperature-robust and weaker on general knowledge than the stable line.
Speculative decoding	The base model's MTP head is not included in these builds.
Pin by digest	Tags are mutable. Pin the artifact digest for production and evaluations.