Qwythos-9B-Claude-Mythos-5-1M
Qwythos-9B is a 9.4B-parameter reasoning model fine-tuned on Qwen3.5-9B with 500M+ tokens of Claude-style reasoning traces. It offers a 1M-token context window (via YaRN scaling), native function calling for tool use, and claimed significant improvements over its base model (+34 MMLU, +30 gsm8k-strict). Marketed as 'uncensored' for sensitive technical domains (cybersecurity, biology, pharmacology). Apache-2.0 licensed, ungated, and available on HuggingFace.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | empero-ai |
| Parameters | 9.4B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 152.5k |
| Likes | 723 |
| Last updated | 2026-06-28 |
| Source | empero-ai/Qwythos-9B-Claude-Mythos-5-1M |
What Qwythos-9B-Claude-Mythos-5-1M is
A full-parameter fine-tune of Qwen/Qwen3.5-9B using supervised fine-tuning (SFT) on internal Claude Mythos/Fable reasoning traces. Ships with YaRN rope-scaling (factor=4.0) baked into config.json, extending native 262k context to 1.048M tokens. Supports OpenAI/Qwen3.5-style function calling without wrapper or tool-specific fine-tune. Evaluation includes lm-evaluation-harness benchmarks (MMLU, gsm8k, ARC, GPQA); tool-use validation on 7 test prompts spanning math, cybersecurity, clinical pharmacology, and biochemistry. No published training code, hyperparameters, or convergence details.
Run Qwythos-9B-Claude-Mythos-5-1M locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="empero-ai/Qwythos-9B-Claude-Mythos-5-1M")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
VRAM: Estimated 18–24 GB (bfloat16) for full model on single GPU. At 1M context, 256k–512k tokens fit on H100/H200 with tensor parallelism or KV-cache offload; full 1M requires multi-GPU or aggressive quantization (ESTIMATE — not validated by card). FP8/INT8 quantization likely reduces to 10–14 GB. Inference batch-1 at 262k native context: ~12–16 GB estimated.
Full-parameter SFT applied by developers; no explicit LoRA/QLoRA guidance provided. Card notes 'full-fine-tune' as applied method. LoRA fine-tuning feasible but untested by authors; QLoRA 4-bit quantization likely viable for modest compute environments. Rope-scaling config in config.json may require caution if extending further; no guidance on gradient checkpointing or mixed-precision tuning stability at 1M context.
When to avoid it — and what to weigh
- Strict content-safety or alignment requirements — Model is intentionally 'uncensored' and designed to engage seriously with red-teaming, exploits, and sensitive medical/pharmaceutical content. Not suitable for applications requiring strict output filtering or safety-alignment guarantees.
- Short-context, latency-critical inference — YaRN scaling trades small short-context quality degradation for long-context performance (acknowledged in card). Single-GPU inference at 1M context requires multi-GPU or KV-cache offload; sub-256k contexts incur memory/latency overhead vs. pure 262k native window.
- Production deployment without risk tolerance — Model is hosted by a small org (Empero AI); no SLA, formal support, or guaranteed maintenance roadmap stated. Evals provided are limited (100-sample per-task harness) and benchmarks are cherry-picked by developers. No third-party independent validation.
- Fact-critical applications without tool grounding — Closed-book reasoning shows mixed results (e.g., −5.0% on GPQA vs. base). Tool-use validation limited to 7 hand-selected prompts. Production pipelines require mandatory retrieval/tool integration to avoid hallucinations on factual queries.
License & commercial use
Apache License 2.0 (OSI-approved). Covers source weights, architecture, and config.json modifications (YaRN scaling). Base model (Qwen/Qwen3.5-9B) is also Apache-2.0. License is permissive and unambiguous.
Apache-2.0 is an OSI-approved permissive license and explicitly permits commercial use, redistribution, and derivative works without royalty or attribution requirement (though attribution is recommended). No gating, no restricted-use terms in card. Commercial deployment, proprietary fine-tuning, and resale are legally permitted. However, no commercial support, SLA, or indemnification offered by Empero AI; buyer assumes liability for model behavior, hallucinations, and uncensored output in production environments.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Strong |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Good |
| Assessment confidence | High |
Model is intentionally uncensored and designed to provide substantive answers on cybersecurity exploitation, red-teaming methodologies, and sensitive biomedical content without refusal. Production use in customer-facing or moderated environments requires robust input validation, output filtering, and clear user disclaimers. No formal security audit, adversarial robustness testing, or jailbreak-resistance evaluation published. Tool-use feature (web search, code execution) introduces dependency on external service availability and code-injection risk if tool execution is not sandboxed. Inference with user-supplied code (Python executor tool) requires trusted environment.
Alternatives to consider
Qwen/Qwen3.5-9B (base model)
Same architecture, aligned variant. Larger safety margin but ~30–34 points lower on reasoning benchmarks. Consider if uncensored behavior is a liability.
Meta Llama-3.1-8B or Llama-3.3-70B
Well-maintained, broader adoption, strong community support, and formal commercial backing. Llama-3.1-8B is comparable in parameter count; Llama-3.3-70B offers stronger reasoning. Trade-off: Llama models are aligned and have stricter refusal patterns; smaller native context (8k–128k).
Mistral Large or Mixtral 8x7B
Permissive licenses, strong reasoning performance, active maintenance. Mistral Large supports longer context natively. Mixtral offers conditional computation. Neither advertises 'uncensored' posture; stricter on sensitive content.
Ship Qwythos-9B-Claude-Mythos-5-1M with senior software developers
Qwythos is a strong fit for long-context reasoning, agentic tool use, and specialized technical domains (cybersecurity, biomedical). If your team requires uncensored substantive Q&A without alignment overhead, run a proof-of-concept with vLLM or SGLang. Validate hallucination rates and tool-selection reliability on your domain before production. Contact Empero AI for scalable deployment guidance.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
Qwythos-9B-Claude-Mythos-5-1M FAQ
Can I use Qwythos in a commercial product without paying licensing fees?
What GPU does Qwythos require?
Does Qwythos support function calling out of the box?
Will Qwythos refuse harmful requests?
Software development & web development with DEV.co
Need help beyond evaluating Qwythos-9B-Claude-Mythos-5-1M? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.
Evaluate Qwythos for Your Use Case
Qwythos is a strong fit for long-context reasoning, agentic tool use, and specialized technical domains (cybersecurity, biomedical). If your team requires uncensored substantive Q&A without alignment overhead, run a proof-of-concept with vLLM or SGLang. Validate hallucination rates and tool-selection reliability on your domain before production. Contact Empero AI for scalable deployment guidance.