Ganymede A1 is the high-capacity member of the Kozu reasoning family. In matched open-harness evaluation, it improves five of seven reported capability rows. The headline GSM8K result reaches 0.913 with 4.7× fewer thinking tokens.

The family principle is intelligence per token: spend reasoning only where it improves the answer. Ganymede A1 is the 27B-class choice for higher-capacity work; Deimos R1 is the compact option and Europa B1 the balanced option. These are deployment roles, not cross-model benchmark claims.

4.7×fewer GSM8K think tokens

+0.449MMLU-Pro absolute delta

6.6×held-out reasoning gain · n=200

Evaluation

Each row compares matched runs in the same open harness. The score table is the source of truth; the token column makes the cost of each result visible. Run-local results should not be compared with vendor-published scores from different evaluation stacks.

Benchmark	Reference	Ganymede A1	Delta	Think tokens · ref → A1
GSM8K · flexible	0.760	0.913	+0.153	1,063 → 226
MMLU-Pro	0.114	0.563	+0.449	156 → 516
IFEval · prompt loose	0.273	0.293	+0.020	2,483 → 2,487
IFEval · instruction loose	0.450	0.454	+0.004	—
IFEval · prompt strict	0.260	0.267	+0.007	—
IFEval · instruction strict	0.437	0.433	−0.004	—
GSM8K · strict `####`	0.767	0.213	−0.554	—

lm-eval 0.4.12 · chat template · few-shot as multi-turn · temperature 1.0 · top_p 0.95 · presence penalty 1.5 · thinking enabled · seed 42 · GSM8K/IFEval n=150 · MMLU-Pro n=25 per subtask.

GSM8K and IFEval used a 16,384-token cap for Ganymede A1 and a 32,768-token cap for the reference; the cap can only constrain Ganymede A1. MMLU-Pro used a symmetric 32,768-token cap.
The MMLU-Pro reference run produced a low score across two independent runs (0.114 / 0.054) and short mean thinking. This is a harness interaction, not a claim about the reference model’s general knowledge.
The held-out, contamination-filtered math audit with verified ground truth moved from 0.065 to 0.430 at n=200. The 12-check instruction probe moved from 1.000 to 0.917 and is directional only.
Strict GSM8K rewards silently copying the few-shot #### N ending. Explicitly requested formats are more reliable; request the format in words.
Raw evaluation artifacts are retained under bench/.

Data & attribution

kozu_reasoning_v1.1 is a roughly 10k-example blend of verified reasoning traces and human-authored instruction data, including a dedicated 800-example format-adherence slice.

Reasoning · ~5kKozu’s Kuiper trace inverter was designed to recreate Mythos/Fable-style reasoning from answer-scrubbed human reference solutions in GSM8K and NuminaMath 1.5. Retained traces passed symbolic answer verification, derivation checks, deterministic defect scans, and an adversarial judge panel.

Instruction · ~5kHuman-authored pairs from Databricks Dolly 15k and OpenAssistant OASST2.

Format · 800Deterministic rewraps of verified rows for explicitly requested formats, including JSON, boxed answers, and named answer markers.

Source licenses

GSM8K: MIT. NuminaMath 1.5: Apache-2.0. Databricks Dolly 15k: CC-BY-SA-3.0. OpenAssistant OASST2: Apache-2.0. Reasoning traces were generated with Kuiper-R1 by Kozu AI.

Deployment

The repository includes merged BF16 weights, tokenizer assets, processor configuration, generation configuration, and the chat template. Thinking appears inside <think>…</think> before the final answer.

vllm serve Michael-Kozu/Ganymede-A1 --served-model-name ganymede-a1 \
  --max-model-len 8192 --gpu-memory-utilization 0.85 --trust-remote-code

Recommended sampling: temperature 0.6–0.8, top_p 0.95, and repetition_penalty about 1.05. Healthy responses are typically well under 1k thinking tokens.

Limitations

This is a research release, not a safety-certified or production-guaranteed system.
At temperature 1.0 with presence penalty, roughly 2–3% of prompts can produce degenerate long generations toward the cap. Use the recommended sampling settings or a max-token cap of at least 8k.
Reasoning is habitually short and does not reliably scale up for extremely hard problems. Coding and tool use are not established by this release.
Narrow constraints such as letter avoidance succeed only about 60–80% in internal probing. Verify outputs for consequential decisions and domain-specific use.