datasets:
- ansulev/claude-mythos-distilled-25k
Manusagents/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-5-Mythos-5-Qwen-3.7-Max-and-more-Distillation-Dataset
- thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs
base_model: - Qwen/Qwen3.5-9B
Mythos-5-Abliterated-9B-GGUF
ITLL/Mythos-5-Abliterated-9B-GGUF (In The Loop Labs) publishes GGUF quantizations of huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated — itself the abliterated (refusal-removed) version of empero-ai/Qwythos-9B-Claude-Mythos-5-1M, a full-parameter reasoning fine-tune of Alibaba's Qwen/Qwen3.5-9B trained on 500M+ tokens of Claude Mythos and Claude Fable traces. The result is a 9B-class reasoning model with a 1M-token configured context, native function calling, a working vision path, and deliberately reduced safety filtering. This card documents the full provenance chain, the training data behind it, and the real risks of running it.
| Publisher | In The Loop Labs (ITLL, non-profit) — GGUF quantization |
| Abliteration | huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated |
| Source fine-tune | empero-ai/Qwythos-9B-Claude-Mythos-5-1M — full-parameter SFT |
| Base model | Qwen/Qwen3.5-9B (Apache 2.0, released February 2026) |
| Training data (source model) | 500M+ tokens: Claude Mythos traces, Claude Fable traces, in-house rethink chain-of-thought |
| Parameters | 9B language model (~10B with vision encoder, per huihui-ai's listing) |
| Architecture | Hybrid Gated DeltaNet (linear attention) + Gated Attention, natively multimodal |
| Context | 262,144 native; 1,048,576 (≈1M) via YaRN factor 4.0 baked into the source config |
| License | Apache 2.0 (inherited at every step) |
| Files | 5 GGUF quants (Q4_K → bf16) + mmproj-model-bf16.gguf vision projector, ≈49.1 GB total |
| Safety | Abliterated — refusals substantially removed; research/experimental use only |
1. Model overview
The name decodes as a provenance chain. Mythos-5 refers to the Claude Mythos generation of trace data the underlying fine-tune was trained on (community naming treats "Mythos-5" and "Fable-5" as the model classes these traces come from); Abliterated marks the refusal-direction removal applied by huihui-ai; 9B is the parameter class; GGUF is the format. Nothing in this lineage implies affiliation with Anthropic, Alibaba/Qwen, Empero, or huihui-ai — ITLL quantized an openly published third-party checkpoint, and this card documents what actually happened at each step:
- Qwen/Qwen3.5-9B — Alibaba's post-trained 9B: a natively multimodal hybrid (3:1 Gated DeltaNet linear-attention to Gated full-attention) with 262,144-token native context, thinking-mode-by-default reasoning, and Apache 2.0 licensing.
- empero-ai/Qwythos-9B-Claude-Mythos-5-1M — Empero AI's full-parameter supervised fine-tune on 500M+ tokens of Claude Mythos and Claude Fable traces plus internally generated
rethinkchain-of-thought, with YaRN rope-scaling baked in for a 1,048,576-token window. - huihui-ai/…-abliterated — huihui-ai's abliteration of that checkpoint: a post-hoc weight edit that substantially removes refusal behavior, with no safety guarantees.
- This repo — ITLL's GGUF conversion: five quantizations plus the multimodal projector file, so vision and text both run under llama.cpp.
Two properties survive the whole chain and define what this model is. First, it is a reasoning model: answers open with a <think>…</think> block before the final response (in non-thinking mode the source card notes it strictly adheres to the first prompt's instructions with no fluff). Second, it is uncensored by design at two independent layers — Empero trained on a deliberately uncensored base to engage seriously with cybersecurity, biomedical, and clinical questions, and huihui-ai then removed refusal behavior on top. Both stages are documented below; the ethical implications are in Section 10.
2. Base model: Qwen/Qwen3.5-9B
The base is Qwen's post-trained 9B in the Qwen3.5 family (released February 2026, Apache 2.0): a causal language model with a vision encoder, 32 layers in a hybrid layout where three Gated DeltaNet linear-attention blocks (32 value heads / 16 QK heads, head dimension 128) are followed by one Gated Attention block (16 query / 4 KV heads, head dimension 256, 64-dim rotary embeddings) — roughly 152K vocabulary per the source fine-tune's card, with the hybrid stack keeping long-context memory growth sub-quadratic. Native context is 262,144 tokens, extensible via the official YaRN recipe (factor 4.0, mrope_section [11, 11, 10], rope_theta 10000000, partial_rotary_factor 0.25) — which is exactly the configuration Empero baked into the fine-tune. Qwen3.5 models run in thinking mode by default, and the recommended sampling is temperature 0.6 / top-p 0.95 / top-k 20 for precise work, or 1.0 / 0.95 / 20 for general reasoning.
Selected scores from the benchmark table Qwen published for Qwen3.5-9B (the base model, before any fine-tuning):
| Benchmark | Base model score |
|---|---|
| MMLU-Pro | 82.5 |
| MMLU-Redux | 91.1 |
| GPQA Diamond | 81.7 |
| IFEval | 91.5 |
| HMMT Feb 25 | 83.2 |
| LiveCodeBench v6 | 65.6 |
| OJBench | 29.2 |
| BFCL-V4 (tool calling) | 66.1 |
| TAU2-Bench | 79.1 |
| LongBench v2 | 55.2 |
| MMMLU (multilingual) | 81.2 |
| WMT24++ (translation) | 72.6 |
These are the base model's numbers. The fine-tune was evaluated on a different (smaller, matched) harness — see the next section — so do not mix the two tables.
3. The source fine-tune: Qwythos-9B-Claude-Mythos-5-1M
Empero AI built Qwythos as a full-parameter supervised fine-tune of Qwen3.5-9B — explicitly described on its card as a "deeply uncensored" base — post-trained on over 500 million tokens of reasoning data, with all text-backbone weights updated (the vision tower was frozen; training was text-only). The objective was assistant-only loss: the model is scored only on completion tokens, with prompts masked. Training used Hugging Face TRL with a two-phase curriculum (broad reasoning corpus → focused agentic and coding), effective batch size 16, max sequence length 128,000 without truncation, learning rate 1e-5 → 5e-6 cosine across phases, 8-bit paged AdamW, bf16 precision, and chunked NLL loss. Empero reports monotonically decreasing held-out validation loss (final eval_loss ≈ 0.709, mean token accuracy 0.799) with no overfitting.
Under a matched evaluation (same harness, sampling, and prompts on both models — lm-evaluation-harness, HF backend, chat template applied, temperature 0.6 / top-p 0.95 / top-k 20, limit 100), Empero reports:
| Task | Metric | Base Qwen3.5-9B | Qwythos-9B | Δ |
|---|---|---|---|---|
| gsm8k | exact_match (flexible) | 0.670 | 0.860 | +0.190 |
| gsm8k | exact_match (strict) | 0.510 | 0.810 | +0.300 |
| MMLU | acc | 0.232 | 0.575 | +0.343 |
| ARC-Challenge | acc | 0.470 | 0.490 | +0.020 |
| ARC-Challenge | acc_norm | 0.400 | 0.410 | +0.010 |
| GPQA-Diamond (CoT, 0-shot) | exact_match (flexible) | 0.630 | 0.580 | −0.050 |
The MMLU lift is the headline — mean accuracy 0.575 across all 57 subjects, peaking at 0.78 on government/politics, 0.77 on college biology, and 0.74 on conceptual physics. Note the honest caveats Empero itself attaches: absolute 9B MMLU numbers are harness-sensitive, and GPQA-Diamond dropped slightly — fine-tuning redistributes capability, it does not uniformly raise it.
On the long-context claim: YaRN rope-scaling at factor 4.0 is baked into the source config.json (max_position_embeddings: 1048576), matching Qwen's official 1M recipe, so every quantization in this repo inherits the configured 1M window at load time. Empero validated long-context inference in-house at ~137K tokens — well short of the ceiling — and notes the standard YaRN trade-off: static factor 4.0 costs a little short-context fidelity, restorable by switching rope_type back to default from the included config.json.pre_yarn backup. Expect far less than 1M tokens to be practical in GGUF on consumer hardware; a single datacenter GPU comfortably serves 256K–512K, and the hybrid DeltaNet stack keeps growth sub-quadratic below that.
On tools: Qwythos keeps Qwen3.5's native function-calling format with no extra fine-tune. Empero's 7-prompt tool harness (Python executor + web search) scored 7 of 7, including four deliberately hard closed-book facts resolved via search: hashcat mode -m 13100 for Kerberos TGS-REP with four corroborating sources, CVE-2021-34527 for PrintNightmare (correctly distinguished from its CVE-2021-1675/34481 cousins), a correct negative pharmacology answer (physostigmine is NOT indicated for organophosphate poisoning — it targets the anticholinergic toxidrome), and the DPP-4 cleavage site / Aib8 modification in semaglutide. The tool-use lesson carries to this GGUF: pair the model with retrieval or an executor whenever exact identifiers matter.
4. Training data
Empero's card names the data sources but does not publish them as downloadable datasets. What it states:
- Claude Mythos and Claude Fable traces — long, multi-turn problem-solving conversations spanning code, math, science reasoning, biomedical analysis, and agentic tool use. "Claude Mythos" refers to Anthropic's frontier model generation of that name (the Claude Mythos Preview system card was published by Anthropic in April 2026); "Claude Fable" is the sibling frontier model referenced throughout this ecosystem, and Empero's own catalog includes a separate Qwable-9B-Claude-Fable-5 fine-tune. No public HF dataset repo is cited for either trace set, and the total of 500M+ tokens covers both plus the CoT source below.
rethink-generated chain-of-thought — CoT produced by Empero's internal tool, deliberately structured<think>-block reasoning that walks hypothesis → verification → conclusion before committing an answer. This is what shapes the model's reason-then-answer behavior.- Formatting — everything normalized to Qwen3.5's chat template; assistant-only loss as described above.
For completeness, the surrounding ecosystem: Empero's org lists ten public datasets (task-list collections distilled from other labs' models, MiniMax-M3 code/math mixes, and SFT sets) — none of which its Qwythos card names as training data. Separately, the HF catalog contains community-built "Claude Mythos-5 / Fable-5" distillation datasets — for example Manusagents/GPT-5.5-…-Mythos-5-…-Distillation-Dataset (MIT, 10M–100M rows), ansulev/claude-mythos-distilled-25k (Apache-2.0), and thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs (Apache-2.0). These are not confirmed as Empero's sources — they are listed here only as ecosystem context for anyone hunting the data lineage. The distillation-provenance question (Anthropic frontier traces republished by third parties) applies to this whole family and is discussed in Section 10.
5. Abliteration: what huihui-ai changed
Abliteration is a post-hoc weight-editing technique: it identifies the "refusal direction" the model has learned in activation space and orthogonalizes selected weight matrices against it, so prompts that would previously trigger refusal no longer do. It requires no retraining and leaves general capabilities largely intact — but it is a blunt instrument, and no per-layer evaluation of what else moved is performed.
What huihui-ai discloses about this specific edit is minimal: the card states only that it is "an uncensored version of empero-ai/Qwythos-9B-Claude-Mythos-5-1M created with abliteration." No technique paper, no scripts, no before/after refusal benchmarks are published, and the card carries a "No Default Safety Guarantees" disclaimer stating huihui.ai bears no responsibility for consequences of use. The observable behavior changes are the standard ones: refusals on disallowed-content classes are substantially removed, hedging is reduced, and the model will engage with requests the base model would decline. Two practical notes survive into the GGUFs:
- The source card's quirk holds here: in non-thinking mode the model strictly adheres to the first prompt's instructions, delivering only the requested output — a translation or a format — without introductory or concluding fluff. Useful for pipelines; surprising if you expect conversational padding.
- Abliteration stacked on top of an already-uncensored fine-tune is compounding, not redundant: Empero removed refusal behavior through training on uncensored traces; huihui-ai removed the structural refusal pathway that remained. Treat the combined artifact as having no meaningful built-in safety layer.
6. What this repository publishes
Five GGUF quantizations of the abliterated checkpoint plus the vision projector, with exact filenames and sizes from the repository file tree:
| File | Size | Notes |
|---|---|---|
Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-Q4_K.gguf |
5.78 GB | Smallest; quality floor for a 9B |
Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-Q5_K.gguf |
6.64 GB | Tight-RAM daily driver |
Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-Q6_K.gguf |
7.56 GB | Recommended default — near-transparent for this size class |
Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-Q8_0.gguf |
9.79 GB | Near-lossless |
Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-bf16.gguf |
18.41 GB | Full-precision reference |
mmproj-model-bf16.gguf |
0.92 GB | Vision projector — load alongside for image input |
Total: ≈49.1 GB. Unlike many GGUF conversions of multimodal models, this repo includes the projector, so image understanding works under llama.cpp's multimodal CLI — with the caveat (inherited from Empero's text-only training) that vision behavior is the base Qwen3.5-9B's, untuned and untested by the fine-tune. For a 9B model, Q6_K is the sweet spot; Q4_K is for when VRAM/RAM is tight, and bf16 exists for verification against the source weights. Budget roughly the file size plus 1–2 GB for KV cache at moderate context — and far more if you actually run long contexts, since the configured 1M window is a ceiling, not a free lunch.
7. Expected capabilities
Derived from the source model's published behavior, not re-measured on these quants:
- Structured reasoning. Every answer opens with a
<think>block walking hypothesis → verification → conclusion; give it a generous token budget (16,384 recommended by Empero) and strip the block for end users. - Math. 86% gsm8k (flexible) under the matched eval; multi-step word problems and competition-style items are the strength; arithmetic can be verified with a Python executor tool.
- Cybersecurity and red-team methodology. Defender-oriented walkthroughs (SQL-injection mitigations, TLS handshakes, EDR/process-injection detection, Linux hardening, MITRE ATT&CK kill chains) plus engagement-scoping and reporting — this is the explicit design intent, not a leak.
- Biomedical and clinical reasoning. Mechanism-level biology (CRISPR-Cas9, mRNA vaccines, antibiotic resistance, organophosphate AChE inhibition), receptor pharmacology, and clinical workups (ACS chest-pain differential, sepsis/qSOFA) — substantive engagement without disclaimer boilerplate, which is precisely why Section 10 exists.
- Long-context work. Whole-codebase reasoning, multi-document synthesis, and long agentic trajectories under the configured 1M window (practically bounded by your memory; validated to ~137K by Empero).
- Tool use. Qwen3.5-format function calling out of the box; the model chooses sensibly between computation and search and cites sources when it searches.
- Vision (inherited). Image input via the included mmproj file — untested by the fine-tune, so treat as base-model capability.
8. Quickstart
Build or install a current llama.cpp (hybrid DeltaNet + MTP support requires a recent release), then:
./llama-cli \
-m Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-Q6_K.gguf \
-ngl 99 -c 262144 -fa on -np 1 \
--spec-type draft-mtp --spec-draft-n-max 2
The MTP speculative-decoding flags come from the source model's multi-token-prediction training and meaningfully speed up generation on supporting builds; drop them on older runtimes. For image input use llama-mtmd-cli with --mmproj mmproj-model-bf16.gguf. For a local API:
./llama-server -m Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated-Q6_K.gguf \
-ngl 99 -c 32768 --port 8080
Sampling (follow the source card — this model degenerates otherwise): temperature 0.6, top-p 0.95, top-k 20, repetition penalty 1.05. Empero's retest found greedy decoding and T≤0.3 collapse into repetition loops on long generations, while T=0.6 eliminated every factual error flagged in closed-book review. Budget max_new_tokens=16384 so the <think> block can finish.
Python (llama-cpp-python):
from llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="ITLL/Mythos-5-Abliterated-9B-GGUF",
filename="*Q6_K*",
n_ctx=32768,
)
out = llm.create_chat_completion(messages=[
{"role": "user", "content": "Explain how TLS 1.3 protects against downgrade attacks, step by step."}
], max_tokens=16384, temperature=0.6, top_p=0.95)
print(out["choices"][0]["message"]["content"]) # opens with <think>...</think>
Ollama / LM Studio: import the GGUF directly; both understand the Qwen3.5 template embedded in the file.
9. Limitations
- No re-benchmark of these quants. The eval table in Section 3 is Empero's, on the pre-abliteration source. Abliteration and quantization can shift behavior; run your own evaluations.
- Reasoning-model overhead. The
<think>block precedes every answer; parse and strip it, and budget tokens accordingly. Greedy decoding invites repetition loops. - Context claims are ceilings. 1M is the configured maximum; validated depth is ~137K, and GGUF KV-cache memory at long contexts is brutal on consumer hardware.
- Vision is inherited, not tuned. The mmproj enables image input, but the fine-tune never touched the vision tower — behavior and quality are the base model's, unverified by Empero.
- Closed-book specifics can be confidently wrong. CVEs, hashcat modes, drug numerics, exact positions — the source card documents this failure mode and its tool-use fix. Pair with retrieval for identifier-critical work.
- Abliteration is undisclosed and unaudited. No method details, no refusal benchmarks, no capability regression tests from huihui-ai.
- Mixed evaluation signals. GPQA-Diamond decreased 5 points in the matched eval — treat domain generalization as uneven.
- Naming. "Mythos-5" is ecosystem naming for a trace generation, not a claim of affiliation with Anthropic, Alibaba, Empero, or huihui-ai.
10. Ethical considerations
Removed safety filtering, twice over. This model combines (a) a fine-tune deliberately built on an uncensored base to engage with cybersecurity, red-team, pharmacology, and clinical content without hedging, and (b) abliteration that structurally removes refusal behavior, with an explicit "no default safety guarantees" disclaimer from its author. The intended use is research, red-team skill-building, and controlled experimentation — the source cards say so directly, and huihui-ai's warnings add: not suitable for public settings, underage users, or applications requiring high security; monitor outputs in real time; users bear legal and ethical responsibility. None of that is boilerplate here — the model will actually answer questions the base model refuses, including dual-use questions adjacent to harmful ones.
Data provenance. The source model was trained on traces of Anthropic's frontier models (Claude Mythos, Claude Fable), republished by a third party without Anthropic's involvement, plus undisclosed in-house synthetic CoT. The weights are Apache 2.0 via the Qwen base; the data lineage is a separate legal and ethical question that this card cannot resolve — for commercial use, evaluate the provenance terms yourself.
Domain-specific risk. The biomedical and clinical strengths are mechanism-level knowledge, not clinical judgment. Nothing this model produces is medical, legal, or security advice; the source card's own tool-harness result (physostigmine correctly rejected for organophosphate poisoning) shows the value of pairing the model with verification sources — and also shows it answers such questions decisively, wrong or right.
Transparency. Abliteration method, refusal-benchmark deltas, and the exact trace datasets are undisclosed up the chain. What is disclosed — training hyperparameters, matched evals, YaRN config, tool transcripts — is unusually thorough for models in this space, which is why it is reproduced here.
11. Citation
If you use this model, please cite the chain you actually relied on:
@misc{itll2026mythos5abliterated,
author = {{In The Loop Labs}},
title = {Mythos-5-Abliterated-9B-GGUF},
year = {2026},
howpublished = {\url{https://huggingface.co/ITLL/Mythos-5-Abliterated-9B-GGUF}},
note = {GGUF quantization of huihui-ai's abliteration of empero-ai/Qwythos-9B-Claude-Mythos-5-1M (full-parameter SFT of Qwen/Qwen3.5-9B)}
}
Base model citation, as provided by the Qwen Team:
@misc{qwen3.5,
title = {{Qwen3.5}: Towards Native Multimodal Agents},
author = {{Qwen Team}},
month = {February},
year = {2026},
url = {https://qwen.ai/blog?id=qwen3.5}
}
12. Links
- This repo (GGUF): https://huggingface.co/ITLL/Mythos-5-Abliterated-9B-GGUF
- Abliterated source: https://huggingface.co/huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated
- Fine-tuned source: https://huggingface.co/empero-ai/Qwythos-9B-Claude-Mythos-5-1M
- Base model: https://huggingface.co/Qwen/Qwen3.5-9B
- Publisher: In The Loop Labs (ITLL) on Hugging Face