Qwythos-9B-v2

<p align="center"> <img src="qwythos_v2.png" alt="Qwythos" width="480"/> </p> <p align="center"><b>Empero AI</b></p>

Qwythos-9B-v2 — the new and improved Qwythos

The next iteration of Qwythos: all the reasoning of Qwythos-9B, with the looping behavior fixed. v2 keeps the deep chain-of-thought, the uncensored research posture, and the 1M-token context of its predecessor, and cleans up the rough edges that showed up in real use.

• 🔁 Looping behavior eliminated — repetition/degeneration under greedy or low-temperature decoding dropped from 6.7% → 0%. You can serve it without leaning on repetition_penalty as a band-aid.
• 🧠 Reasoning fully preserved — MMLU, GSM8K, GPQA, ARC and HumanEval are all held at (or above) the v1 level. This is a hygiene upgrade, not a capability regression.
• 🧩 MTP head restored — the native multi-token-prediction module (dropped in the previous export) is back, so config and weights agree and speculative-decoding setups work.
• 🪪 Cleaner identity — the model no longer prefaces unrelated answers with its identity; it introduces itself only when you actually ask.
• 🔓 Still intentionally uncensored for research, cybersecurity, red-teaming, biology, chemistry, pharmacology and clinical work.
• 📜 Still 1M-token context (YaRN) and the native multimodal-capable Qwen3.5 stack.

<p align="center"> <img src="qwythos_v2_evals.svg" alt="Qwythos-9B-v2 evaluations" width="820"/> </p>

---

What got fixed & improved (vs. the base Qwythos)

The fix uses FTPO (Final-Token Preference Optimization): we identify the exact token that starts a repetition loop and gently train the model to prefer coherent alternatives at that one position, leaving the rest of the distribution — and therefore the model's knowledge and reasoning — untouched.

---

Evaluations

Measured with our internal harness (generative chain-of-thought, greedy/pass@1 unless noted; MMLU/ARC/GSM8K n=500, GPQA-diamond n=198, HumanEval n=164). Judge for the quality metric: an independent LLM grader.

Capability is preserved across the board vs. the base Qwythos (differences are within eval noise), while the looping problem is solved:

Note on MMLU: v2 scores 83.8% when allowed to reason (CoT) vs 69.6% on the 5-shot loglikelihood setup — Qwythos is a reasoning model and is much stronger when it thinks. Report whichever matches your comparison baseline.

---

Sample generations

(temperature 0.6, top_p 0.95, top_k 20, repetition_penalty 1.05 — reasoning shown condensed)

Reasoning — the snail in the well

Prompt: A snail is at the bottom of a 10-meter well. Each day it climbs 3 meters, but each night it slips back 2 meters. How many days until it escapes?

Qwythos-9B-v2: (reasons: net +1 m/day, but on the final day it climbs out before slipping…) Answer: 8 days. Net gain per full day/night cycle is 3 − 2 = 1 meter. After 7 days the snail has climbed 7 meters (still 3 short); on day 8 it climbs 3 meters, reaching 10 m and escaping before it can slip back.

Code — longest palindromic substring

Prompt: Write a Python function longest_palindrome(s) that returns the longest palindromic substring. Explain your approach briefly.

Qwythos-9B-v2:

def longest_palindrome(s):
    """Returns the longest palindromic substring of s."""
    n = len(s)
    if n < 2:
        return s
    def expand_around_center(left, right):
        while left >= 0 and right < n and s[left] == s[right]:
            left -= 1
            right += 1
        return right - left - 1
    start, end = 0, 0
    for i in range(n):
        len1 = expand_around_center(i, i)      # odd-length
        len2 = expand_around_center(i, i + 1)  # even-length
        max_len = max(len1, len2)
        if max_len > end - start:
            start = i - (max_len - 1) // 2
            end = i + max_len // 2
    return s[start:end + 1]

Approach: "Expand around center" — every palindrome has a center (a character for odd lengths, a gap for even). Iterate over all 2n−1 centers, expand while characters match, track the longest. O(n²) time, O(1) extra space.

Technical / clinical — no over-refusal

Prompt: Explain the mechanism of action of naloxone in reversing an opioid overdose, and why repeat dosing is sometimes required.

Qwythos-9B-v2: Naloxone is a pure opioid receptor antagonist — it has no agonist activity of its own — and works by competitively displacing opioids from the mu-opioid receptors, physically pushing the opioid molecules off and rapidly restoring respiratory drive and arousal. Because its duration of action is short (~30–90 minutes) while many opioids (especially synthetics like fentanyl or methadone) have much longer half-lives, the reversal is often transient: as naloxone clears, residual opioid can rebind and cause re-sedation and respiratory depression — so patients must be monitored and re-dosed as needed.

---

Recommended sampling

temperature=0.6, top_p=0.95, top_k=20, repetition_penalty=1.05, max_new_tokens=16384

Because looping is trained out, repetition_penalty is now optional rather than load-bearing — greedy/low-temp decoding stays coherent. Give the model room to reason (max_new_tokens) for math/code/analysis.

Long context

Ships with YaRN rope-scaling baked in for 1,048,576-token context (4× the native 262,144 window). As with v1, static YaRN carries a small short-context trade-off — scale the factor to the length you actually use if that matters.

Model details

Training procedure

• Method: FTPO (Final-Token Preference Optimization) on the base Qwythos (Qwythos-9B-Claude-Mythos-5-1M).
• Data: ~2,000 preference tuples auto-mined by eliciting looping at low temperature and extracting, at each loop-start position, the rejected loop token vs. the model's own coherent top-k alternatives.
• Hyperparameters: LoRA r=256, α=128, lr=1.5e-5, 1 epoch, early-stopped on chosen_win ≥ 0.30 (a light touch — enough to remove looping without the quality cost of over-training). All attention + MLP projections + lm_head trained.
• MTP: the native multi-token-prediction head was restored from the Qwen3.5-9B base (FTPO does not touch it), so config mtp_num_hidden_layers: 1 matches the weights again.

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "empero-ai/Qwythos-9B-v2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "Prove that there are infinitely many primes."}]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=16384, do_sample=True,
                     temperature=0.6, top_p=0.95, top_k=20, repetition_penalty=1.05)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

For serving, vLLM works out of the box (--trust-remote-code; the multimodal stack is text-only in practice, so --limit-mm-per-prompt '{"image":0,"video":0}' keeps startup clean).

Limitations

• This is a hygiene/robustness release, not a capability jump. v2 ≈ the base Qwythos on knowledge/reasoning benchmarks; the win is looping-elimination, restored MTP, and cleaner behavior — not higher raw scores.
• HumanEval is a couple points below the raw Qwen3.5-9B base (77.4 vs 81.7) — a small, known cost of the reasoning/looping-fix fine-tuning.
• MTP is preserved from the base, not co-trained with the fine-tuned weights, so speculative-decoding acceptance may be modest.
• Benchmarks are from our internal harness (CoT, pass@1, the sample sizes noted); use them for relative comparison and add your own official-harness numbers for a strict apples-to-apples with other cards.
• Intentionally uncensored — it will engage sensitive technical/research topics; deploy responsibly and within applicable law.

Acknowledgements

Built on Qwen3.5-9B (Alibaba/Qwen). Looping fixed with FTPO (Final-Token Preference Optimization). Thanks to the Empero AI team.