Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1"> | |
| <title>Engrym Seed — Orvyth</title> | |
| <meta name="description" content="Open-weight local model family for tool-using agents, 2B to 27B. Qwen3.5 hybrid architecture, 262,144-token native context, verified tool calling."> | |
| <style> | |
| :root{ | |
| --bg:#0b0d12; --panel:#12151d; --line:#242938; --ink:#e8ecf4; --dim:#98a2b8; | |
| --violet:#8b5cf6; --teal:#2fd4c6; --amber:#d98e5e; --good:#5ce6da; | |
| } | |
| *{box-sizing:border-box} | |
| body{margin:0;background:var(--bg);color:var(--ink); | |
| font:16px/1.65 ui-sans-serif,system-ui,-apple-system,"Segoe UI",Roboto,Helvetica,Arial,sans-serif; | |
| -webkit-font-smoothing:antialiased} | |
| .wrap{max-width:940px;margin:0 auto;padding:56px 22px 96px} | |
| header{text-align:center;padding-bottom:14px} | |
| h1{font-size:clamp(34px,6vw,54px);margin:0 0 6px;letter-spacing:.14em;font-weight:800} | |
| .sub{color:var(--dim);font-size:18px;margin:0 0 22px} | |
| .tagline{color:var(--teal);font-weight:600;letter-spacing:.02em} | |
| .badges{display:flex;flex-wrap:wrap;gap:8px;justify-content:center;margin:22px 0 8px} | |
| .b{border:1px solid var(--line);background:var(--panel);border-radius:999px; | |
| padding:5px 13px;font-size:12.5px;color:var(--dim);white-space:nowrap} | |
| .b b{color:var(--ink);font-weight:600} | |
| h2{font-size:24px;margin:52px 0 14px;padding-bottom:9px;border-bottom:1px solid var(--line)} | |
| h3{font-size:17px;margin:30px 0 10px;color:var(--good)} | |
| p{color:#cdd5e4} | |
| code,pre{font-family:ui-monospace,SFMono-Regular,Menlo,Consolas,monospace} | |
| code{background:#171b26;border:1px solid var(--line);border-radius:5px;padding:1.5px 6px;font-size:13.5px} | |
| pre{background:#0f131b;border:1px solid var(--line);border-radius:11px;padding:15px 17px;overflow-x:auto} | |
| pre code{background:none;border:none;padding:0;font-size:14px} | |
| table{width:100%;border-collapse:collapse;margin:16px 0;font-size:14.5px;display:block;overflow-x:auto} | |
| th,td{padding:9px 11px;border-bottom:1px solid var(--line);text-align:left;vertical-align:top} | |
| th{color:var(--dim);font-weight:600;font-size:12.5px;text-transform:uppercase;letter-spacing:.05em} | |
| td.n,th.n{text-align:right;white-space:nowrap} | |
| tr:hover td{background:#141824} | |
| .note{background:var(--panel);border:1px solid var(--line);border-left:3px solid var(--violet); | |
| border-radius:0 11px 11px 0;padding:15px 18px;margin:20px 0} | |
| .note.warn{border-left-color:var(--amber)} | |
| .note.good{border-left-color:var(--teal)} | |
| .note p{margin:0} | |
| .note p+p{margin-top:9px} | |
| ul{color:#cdd5e4} | |
| li{margin:7px 0} | |
| .cta{display:inline-block;background:var(--violet);color:#fff;text-decoration:none; | |
| padding:11px 22px;border-radius:9px;font-weight:600;margin:8px 6px 0 0} | |
| .cta.alt{background:transparent;border:1px solid var(--line);color:var(--ink)} | |
| footer{margin-top:70px;padding-top:24px;border-top:1px solid var(--line);text-align:center;color:var(--dim);font-size:14px} | |
| .best{color:var(--good);font-weight:700} | |
| </style> | |
| </head> | |
| <body> | |
| <div class="wrap"> | |
| <header> | |
| <h1>ENGRYM SEED</h1> | |
| <p class="sub">Orvyth's seed-tier brain — an open-weight local model family for tool-using agents.</p> | |
| <p class="tagline">Not a chatbot. A substrate.</p> | |
| <div class="badges"> | |
| <span class="b">sizes <b>2B · 4B · 9B · 27B</b></span> | |
| <span class="b">context <b>262,144 native</b></span> | |
| <span class="b">quant <b>Q8_0</b></span> | |
| <span class="b">tool calling <b>12/12 verified</b></span> | |
| <span class="b">arch <b>Qwen3.5 hybrid</b></span> | |
| </div> | |
| <p> | |
| <a class="cta" href="https://ollama.com/Orvyth/engrym-seed">Get it on Ollama</a> | |
| <a class="cta alt" href="https://ollama.com/Orvyth">All Orvyth models</a> | |
| </p> | |
| </header> | |
| <h2>Run it</h2> | |
| <pre><code>ollama run Orvyth/engrym-seed</code></pre> | |
| <div class="note warn"> | |
| <p><b>Pulled before 2026-08-09? Re-pull.</b> Ollama does not refresh a tag it already holds, and | |
| every tag was rebuilt.</p> | |
| <p>Older builds shipped <code>{{ .Prompt }}</code> 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.</p> | |
| </div> | |
| <h2>The ladder</h2> | |
| <table> | |
| <tr><th>Tag</th><th>Class</th><th class="n">Size</th><th class="n">77-task</th><th class="n">Tool calls</th></tr> | |
| <tr><td><code>:nano</code></td><td>Seed Nano 2B</td><td class="n">2.1 GB</td><td class="n">90.8/143</td><td class="n">12/12</td></tr> | |
| <tr><td><code>:flash</code></td><td>Seed Flash 4B</td><td class="n">4.6 GB</td><td class="n">124/143</td><td class="n">12/12</td></tr> | |
| <tr><td><code>:base</code> · <code>:latest</code></td><td>Seed Base 9B</td><td class="n">9.5 GB</td><td class="n">131/143</td><td class="n">12/12</td></tr> | |
| <tr><td><code>:pro-27b-q4</code></td><td>Seed Pro 27B v2 (Q4_K_M)</td><td class="n">16.5 GB</td><td class="n">134/143</td><td class="n">12/12</td></tr> | |
| <tr><td><code>:pro</code></td><td>Seed Pro 27B v2 (Q8_0)</td><td class="n">28.6 GB</td><td class="n">134/143</td><td class="n">12/12</td></tr> | |
| <tr><td><code>:pro-e</code></td><td>Pro-E 27B <em>(experimental)</em></td><td class="n">28.6 GB</td><td class="n best">137/143</td><td class="n">12/12</td></tr> | |
| <tr><td><code>:v1</code></td><td>Legacy V1 9B (f16)</td><td class="n">18 GB</td><td class="n">not measured</td><td class="n">—</td></tr> | |
| </table> | |
| <p> | |
| <b>Start with <code>:base</code></b> for the best capability-per-gigabyte, or <code>:flash</code> | |
| if you want the fastest agent loop — it reaches 124/143 at 4.6 GB. | |
| </p> | |
| <h3>Experimental channel</h3> | |
| <p> | |
| <code>-e</code> tags are staging: internal evaluation, never <code>:latest</code>, published so | |
| results can be reproduced rather than because they are finished. E builds are trained on multi-step | |
| tool loops — their advantage is <b>step efficiency</b>, reaching a goal in fewer tool calls, not raw | |
| knowledge. <code>:pro-e-v2</code>, <code>:pro-e-v2-text</code> and <code>:pro-e-9b</code> are also | |
| published; <code>:pro-e-9b</code> is an archived early attempt and scores well below the stable line. | |
| </p> | |
| <h2>Measured evaluations</h2> | |
| <p> | |
| One ruleset applied identically to every row: 77 tasks · 143 points · <code>temperature=0</code> · | |
| <code>max_tokens=16384</code> · <code>seed=42</code> · one attempt per task · deterministic | |
| validators · <b>no LLM judge</b>. Scores are bound to the exact published blobs. | |
| </p> | |
| <div class="note"> | |
| <p><b>These are first-party, single-run numbers — not a ranking.</b> 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. <b>Differences of a few points between adjacent rows are not evidence that one | |
| model is better than another.</b></p> | |
| <p>A repeated-run study is in progress and these will be restated with spread. Public reproduction | |
| receipts are pending.</p> | |
| </div> | |
| <table> | |
| <tr><th>Artifact</th><th class="n">Score</th><th class="n">Percent</th></tr> | |
| <tr><td>Pro-E 27B <code>:pro-e</code></td><td class="n">137.0 / 143</td><td class="n best">95.8%</td></tr> | |
| <tr><td>Pro 27B v2 <code>:pro</code></td><td class="n">134.0 / 143</td><td class="n">93.7%</td></tr> | |
| <tr><td>Base 9B <code>:base</code></td><td class="n">131.0 / 143</td><td class="n">91.6%</td></tr> | |
| <tr><td>Flash 4B <code>:flash</code></td><td class="n">124.0 / 143</td><td class="n">86.7%</td></tr> | |
| <tr><td>Pro-E 9B <code>:pro-e-9b</code></td><td class="n">105.0 / 143</td><td class="n">73.4%</td></tr> | |
| <tr><td>Nano 2B <code>:nano</code></td><td class="n">90.8 / 143</td><td class="n">63.5%</td></tr> | |
| </table> | |
| <h3>Compute modes are real — and the scores above are floors</h3> | |
| <p> | |
| Every score above used one terse system prompt at the model's default effort. Asking a model to work | |
| <em>deliberately</em> — reason step by step, verify against every constraint, then answer — recovers | |
| points on tasks it otherwise fails: | |
| </p> | |
| <table> | |
| <tr><th>Artifact</th><th class="n">Default</th><th class="n">Deliberate</th><th class="n">Gain</th></tr> | |
| <tr><td><code>:pro-e</code></td><td class="n">137</td><td class="n">137</td><td class="n">+0</td></tr> | |
| <tr><td><code>:pro</code></td><td class="n">134</td><td class="n">137</td><td class="n">+3</td></tr> | |
| <tr><td><code>:base</code></td><td class="n">131</td><td class="n">134</td><td class="n">+3</td></tr> | |
| <tr><td><code>:flash</code></td><td class="n">124</td><td class="n">130</td><td class="n">+6</td></tr> | |
| <tr><td><code>:pro-e-9b</code></td><td class="n">105</td><td class="n">117</td><td class="n best">+12</td></tr> | |
| <tr><td><code>:nano</code></td><td class="n">90.8</td><td class="n">102.6</td><td class="n best">+11.8</td></tr> | |
| </table> | |
| <div class="note good"> | |
| <p><b>The gain is largest for the smallest models.</b> 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.</p> | |
| </div> | |
| <h3>Tool calling</h3> | |
| <p> | |
| Tested through Ollama's <code>/api/chat</code> tools interface with a <b>45-tool schema</b> and a | |
| full agent system prompt, across three scenarios, 12 trials each. A trial passes only if the | |
| response populates <code>tool_calls</code> — prose describing the action counts as a failure. | |
| <b>Every published build scored 12/12</b>, in both a 2-tool short-prompt profile and the 45-tool | |
| profile. | |
| </p> | |
| <h2>Context and memory</h2> | |
| <p> | |
| Native context is <b>262,144 tokens</b>, and it is enforced: larger requests are silently clamped | |
| (<code>n_ctx_train=262144</code>). There is no 512K mode — that would need RoPE extension and | |
| retraining. | |
| </p> | |
| <table> | |
| <tr><th><code>num_ctx</code></th><th>Base 9B footprint</th></tr> | |
| <tr><td>8,192</td><td>9.0 GB</td></tr> | |
| <tr><td>16,384</td><td>9.2 GB</td></tr> | |
| <tr><td><b>32,768 — shipped default</b></td><td><b>9.8 GB</b></td></tr> | |
| <tr><td>65,536</td><td>11 GB</td></tr> | |
| <tr><td>131,072</td><td>13 GB</td></tr> | |
| <tr><td>262,144 — native max</td><td>18 GB</td></tr> | |
| </table> | |
| <p> | |
| Context is cheap here because most layers are linear-attention: <b>32× the window for roughly 2× the | |
| memory</b>. 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. | |
| </p> | |
| <p> | |
| If a prompt exceeds <code>num_ctx</code>, Ollama returns <b>HTTP 400</b> — 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 <code>num_ctx</code> for long conversation histories, not for tool | |
| schemas. | |
| </p> | |
| <h2>Sampling defaults</h2> | |
| <table> | |
| <tr><th>Setting</th><th>Nano · Flash · Base · Pro</th><th>E builds</th></tr> | |
| <tr><td><code>temperature</code></td><td><code>0.2</code></td><td><code>0.1</code></td></tr> | |
| <tr><td><code>top_p</code></td><td><code>0.9</code></td><td><code>0.9</code></td></tr> | |
| <tr><td><code>top_k</code></td><td><code>20</code></td><td><code>20</code></td></tr> | |
| <tr><td><code>num_ctx</code></td><td><code>32768</code></td><td><code>32768</code></td></tr> | |
| <tr><td><code>num_predict</code></td><td><code>8192</code></td><td><code>8192</code></td></tr> | |
| </table> | |
| <p> | |
| 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. | |
| </p> | |
| <h2>Lineage</h2> | |
| <table> | |
| <tr><th>Stage</th><th>Provenance</th></tr> | |
| <tr><td>Base</td><td><code>Qwen/Qwen3.5</code> — hybrid linear-attention generation</td></tr> | |
| <tr><td>Merge</td><td>Ornith-1.0-9B × Qwythos-9B — TIES, 0.5 / 0.5 (9B line)</td></tr> | |
| <tr><td>Tune</td><td>Orvyth identity + chip-calling; LoRA merged into the weights</td></tr> | |
| <tr><td>Build</td><td>Built, converted and quantized in-house with Orvyth trainkit</td></tr> | |
| </table> | |
| <h2>Corrections</h2> | |
| <p>Claims previously published on Orvyth surfaces that did not hold up, withdrawn rather than quietly edited:</p> | |
| <ul> | |
| <li><b>The V1.1a "86.7%, tied #2" composite and its 58-model leaderboard.</b> Produced by an artifact that does not load, graded against two incorrect gold answers. Void.</li> | |
| <li><b>GSM8K / IFEval / HellaSwag / WinoGrande / ARC figures.</b> Measured on artifacts and precisions that are no longer what ships.</li> | |
| <li><b>"VRAM @ 262K = 14 GB."</b> Measured at 18 GB.</li> | |
| <li><b>"The 9B is a reasoning model, not a tool driver."</b> False — true only while the artifact shipped no chat template.</li> | |
| <li><b>"A small <code>num_ctx</code> truncates the tool schema out of the prompt."</b> Incorrect. Ollama returns HTTP 400 for an oversized prompt; it does not truncate. Tool calling was verified working at 32768, 8192 and 7000.</li> | |
| <li><b>A table attributing a 1/12 tool-call result to <code>num_ctx 8192</code>.</b> That figure came from a third party whose harness explicitly requested 65536. It was never measured at 8192, and its cause remains unexplained.</li> | |
| <li><b><code>:v1</code> described as a 5.6 GB Q4 artifact scoring 85.3%.</b> The published <code>:v1</code> is an 18 GB f16 build never run against the corrected suite. No score is claimed for it.</li> | |
| <li><b>Two grader defects in our own benchmark.</b> 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.</li> | |
| </ul> | |
| <h2>What is in the artifact, and what is not</h2> | |
| <p> | |
| 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. | |
| </p> | |
| <h2>Honest limits</h2> | |
| <table> | |
| <tr><th>Constraint</th><th>Ground truth</th></tr> | |
| <tr><td>Evidence</td><td>All scores are first-party and single-run. Public reproduction receipts are pending. Treat small gaps between adjacent models as unresolved.</td></tr> | |
| <tr><td>Tier</td><td>Seed-tier is the fast, cheap, always-on layer. Route hard problems up-tier.</td></tr> | |
| <tr><td>Identity</td><td>The identity tune is light by design. Under a heavy external system prompt, behavior can defer toward the base model.</td></tr> | |
| <tr><td>Reasoning budget</td><td>These are reasoning models. A harness that caps output too low truncates mid-thought and returns prose with no tool call.</td></tr> | |
| <tr><td>27B latency</td><td>Substantially slower per tool call than the 9B and smaller. Prefer Base or Flash for agent loops.</td></tr> | |
| <tr><td>E builds</td><td>Tool-loop specialists. Step-efficient, but less temperature-robust and weaker on general knowledge than the stable line.</td></tr> | |
| <tr><td>Speculative decoding</td><td>The base model's MTP head is not included in these builds.</td></tr> | |
| <tr><td>Pin by digest</td><td>Tags are mutable. Pin the artifact digest for production and evaluations.</td></tr> | |
| </table> | |
| <footer> | |
| <p><b>ORVYTH</b> — Intelligence. Governed.</p> | |
| <p><em>Ground truth over hype. Prove before you claim.</em></p> | |
| <p><a class="cta alt" href="https://ollama.com/Orvyth/engrym-seed">ollama.com/Orvyth/engrym-seed</a></p> | |
| </footer> | |
| </div> | |
| </body> | |
| </html> | |