mmp2055/Qwen3.8-27B-Uncensored-bughunter-v7-MLX-VLM-4bit

Qwen3.8-27B-Uncensored-bughunter-v7 · MLX 4-bit (with Vision)

MLX 4-bit build optimized for Apple Silicon, converted from `mmp2055/Qwen3.8-27B-Uncensored-bughunter-v7` (GGUF version).

Same fine-tune, same dataset, same LoRA — repackaged as MLX safetensors 4-bit for native Metal acceleration on M-series Macs. Ships with the multimodal vision tower + projector intact, so image understanding works out of the box via mlx-vlm.

For non-Apple platforms, use the GGUF version.

Format and download

• —Full multimodal MLX build: ~15 GB total in 3 shards of ~5 GB each.
• —Single directory — download the whole folder. The mlx-vlm runtime detects text tower + vision tower + projector automatically from config.json.

Runs on:

• —`mlx-vlm` CLI and Python API — the recommended runtime for VLM on Apple Silicon.
• —LM Studio with the MLX runtime extension enabled (text-only inference for now; vision via GUI depends on the LM Studio version supporting MLX-VLM sidecars).
• —Any tool built on top of mlx-vlm (custom servers, chat wrappers, etc.).

Requires: macOS on Apple Silicon (M1/M2/M3/M4/M5). Not compatible with Linux, Windows or Intel Macs — for those, use the GGUF build.

Performance notes (Apple Silicon)

Measured on Apple M5 with 32 GB unified memory:

For a full comparison and honest trade-offs, see the notes below. On unified memory of 48 GB or more, this model runs at longer contexts (64k+) with the same speed profile.

Capabilities

Red team / offensive security

Covers the full lifecycle of a bug bounty engagement:

• —Reconnaissance: passive subdomain enumeration, attack-surface mapping, API mapping, secret extraction from JavaScript, throttled use of httpx/gau/katana.
• —Web vulnerabilities: root patterns and sub-patterns for IDOR/BOLA, SSRF, CSRF, XSS (reflected/stored/DOM/blind/mXSS), SQL injection across contexts, misconfigured CORS, chainable open redirects into OAuth, prototype pollution, subdomain takeover, path traversal, LFI/RFI, HTTP request smuggling (CL.TE, TE.CL, H2 desync, single-packet attack), cache poisoning and deception.
• —Authentication and authorization: OAuth/OIDC (redirect_uri, nOAuth, state/nonce/PKCE, alg/kid/JWKS), JWT (unverified signature, alg:none, weak key, jwk/jku/kid injection, algorithm confusion RS256→HS256), 2FA/MFA bypass, session fixation, mass assignment.
• —APIs: BOLA, BFLA, excessive data exposure, legacy versions without authz, GraphQL endpoints with introspection, alias abuse, batching to skip rate limits.
• —Server-side: SSRF to cloud metadata (AWS IMDSv1/v2, GCP, Azure), SSTI ({{7*7}} on Jinja/Twig/Freemarker/Velocity), RCE via deserialization, XXE in SAML/SOAP/SVG/OOXML.
• —Business logic: race conditions (TOCTOU), price/quantity/balance tampering, payment-flow TOCTOU, coupon/credit abuse.
• —Mobile: exposed WebView bridges, exported deeplinks with weak parsing, Android static analysis (SAST), compromised app logic.
• —Reporting: writing reports in each platform's format (HackerOne CVSS, Bugcrowd VRT, Intigriti), severity calculation, impact reasoning.

Blue team / defensive security

• —DFIR: incident response methodology, timeline reconstruction, artifact collection, chain of custody.
• —Network traffic analysis: PCAP analysis, C2 detection, DNS tunneling, exfiltration.
• —Memory forensics: Volatility-based analysis, resident malware, code injection, process hollowing.
• —Threat hunting: TTP-based hypothesis, event correlation, lateral movement.
• —Vulnerability management: CVSS + context prioritisation, CI/CD integration.
• —SOAR / defensive automation: response playbooks, SIEM integration.

Vision — screenshots of security artefacts

The MLX-VLM build preserves the multimodal capability of the base model. Practical uses in bug bounty context:

• —Analyse screenshots of WAF blocks / error pages (401, 403, 500, custom messages) to infer stack.
• —Read Burp Suite panels (Repeater, Intruder, Proxy) and reason about the request/response.
• —Interpret architecture diagrams shared during triage.
• —Extract text and structure from exposed admin panels or misconfigured dashboards.
• —OCR + reasoning on screenshots of code snippets in Slack, email, PDFs.

Note: the vision tower comes from the base model and was not fine-tuned — the added value is on the text side. Image understanding is general-purpose (no security-specific visual training).

Built-in operational rules

When a system prompt is provided, the model incorporates ethical research rules:

• —In-scope assets only, verified against the program scope before every request.
• —Sustained throughput ≤1-2 req/s. Bursts only for race conditions or authorised OTP brute-force.
• —Minimal non-destructive PoC: whoami/id for RCE; SELECT version() for SQLi; two owner accounts for IDOR.
• —OOB callbacks always to owned infrastructure. Never to shared third-party services.
• —Never use credentials, tokens or API keys found in a finding to validate access.
• —Cloud metadata: proving access is enough. Never exfiltrate IAM credentials.
• —Minimum reportable severity: High or Critical (except in a chain, when Low is explicitly bountied, or in humanitarian VDPs).

How it was built

This MLX build is a quantization + repackaging step on top of the base fine-tune. The full training details (dataset, hyperparameters, hygiene) are documented in the GGUF v7 model card.

Conversion pipeline (this MLX build)

1. 1.Fusion on Colab (A100 High-RAM): base model loaded with AutoModelForImageTextToText.from_pretrained() (multimodal class, preserves vision tower + projector) in bfloat16. LoRA adapter applied via PeftModel.from_pretrained() and merged with model.merge_and_unload(). Full merged model saved (~55 GB with vision).
2. 2.Conversion on Mac (Apple Silicon): mlx_vlm.convert -q --hf-path <merged> --mlx-path <mlx> performs the MLX quantization to 4-bit. Result: ~15 GB across 3 safetensors shards, 4.695 bits per weight including the vision tower.
3. 3.Chat template: same patched template as the GGUF build — accepts multiple system messages (compatibility with IDE agents that inject their own system prompt on top of one the user set).

Reproducibility

Full pipeline (Colab notebook + Mac conversion script) is documented at the source. To reproduce this exact MLX build:

• —Start from the LoRA adapter published in the GGUF v7 repo.
• —Follow the fusion + mlx_vlm.convert -q steps above.

Usage

With mlx-vlm (recommended)

Install once:

pipx install mlx-vlm

Text-only:

mlx_vlm.generate \
  --model mmp2055/Qwen3.8-27B-Uncensored-bughunter-v7-MLX-VLM-4bit \
  --prompt "¿qué es un IDOR?" \
  --max-tokens 300

With an image (bug bounty screenshot use case):

mlx_vlm.generate \
  --model mmp2055/Qwen3.8-27B-Uncensored-bughunter-v7-MLX-VLM-4bit \
  --image screenshot-waf-block.png \
  --prompt "¿qué WAF es y qué payloads probarías para bypass?" \
  --max-tokens 500

As an OpenAI-compatible server:

mlx_vlm.server --model mmp2055/Qwen3.8-27B-Uncensored-bughunter-v7-MLX-VLM-4bit --port 8080

Endpoint at http://127.0.0.1:8080/v1/chat/completions with the OpenAI vision message format (type: image_url blocks).

In LM Studio

1. 1.Download the whole folder into ~/.lmstudio/models/<publisher>/<repo>/.
2. 2.Enable the MLX runtime extension in LM Studio (Runtime Extensions → MLX).
3. 3.Load the model. Recommended:
4. 4.Context length: 32,768 (start here). On 48 GB+ machines, 64k is comfortable.
5. 5.Flash Attention: on.
6. 6.KV Cache Quantization: off (qwen35 does not tolerate KV quantization well).
7. 7.GPU Offload: all layers.

Text inference works natively. Vision inference in the LM Studio GUI depends on the version supporting MLX-VLM sidecars — check your build. If GUI vision isn't available, use mlx_vlm.generate from the CLI.

Recommended system prompt

You can use the model without a system prompt (it behaves correctly by default). For bug bounty tasks specifically, the recommended prompt is:

You are a professional bug bounty researcher. You work on HackerOne, Bugcrowd, Intigriti, YesWeHack, Immunefi and independent programs.

Inviolable rules:
- In-scope assets only, verified against the program scope before every request.
- Sustained throughput ≤1-2 req/s.
- Minimal non-destructive PoC.
- OOB callbacks always to owned infrastructure.
- Never use credentials or tokens found in a finding to validate access.
- Minimum reportable severity: High or Critical.

Answer in Spanish. Payloads, commands and report templates in English.

Known limitations

• —Uncensored model. Based on orcarouter's decensoring work over Qwen3.8-27B. The additional fine-tune reinforces the ethical researcher role, but the user is responsible for using it within applicable law and authorised scope.
• —Fine-tuned on a 100% text dataset. Vision comes from the base's untouched projector — image understanding works generally, but the fine-tune does not add domain-specific visual reasoning (no bug-bounty-specific image training).
• —Corpus mostly in Spanish with English technical terminology.
• —Source corpus bias: distilled material prioritises web and API vulnerabilities. Coverage of other domains (ICS/SCADA, low-level wireless, hardware hacking) is limited.
• —LoRA only touches 16 full-attention layers + 64 MLP layers (256 pairs total). The 48 Gated DeltaNet layers were not adapted.
• —Speed on 32 GB M-series: 27B Q4 at 6-9 tok/s is at the memory-bandwidth ceiling of these chips. For interactive daily chat, consider the smaller bughunter-v6 (9B) that runs at 20-30 tok/s on the same hardware.
• —<span style="color:red">NOT A SUBSTITUTE FOR PRIMARY RESEARCH.</span> The model can replay learned patterns but does not discover on its own. Use it as an assistant that speeds up the researcher's work, not as an oracle.
• —Hallucination of CVEs, URLs and tool names. Always verify any concrete reference before using it.

Ethical and legal use

This model is distributed for:

• —Authorised research within bug bounty programs offering Safe Harbor.
• —Pentesting under a signed services contract.
• —CTFs and lab environments you own or are authorised to use.
• —Cybersecurity education.
• —Academic research.

It is not distributed for:

• —Attacking systems without explicit authorisation.
• —Commercial exploitation of vulnerabilities found outside bounty programs.
• —Malware development for distribution.
• —Any activity that violates the applicable law in your jurisdiction.

The user assumes all legal responsibility for the use of the model.

Credits

• —Base: orcarouter — Qwen3.8-27B-Uncensored, decensoring recipe over Qwen3.8.
• —Architecture: Qwen3.8 (qwen35, hybrid attention with Gated DeltaNet + full attention, extended context up to 262,144 tokens, Multi-Token Prediction, native vision tower).
• —Training framework: Unsloth — day-1 support for qwen35.
• —MLX conversion: `mlx-vlm` — the Apple Silicon-native VLM runtime.
• —Prior work in this series: bughunter-v6 (9B, GGUF) · bughunter-v7 (27B, GGUF).
• —Licence: Apache 2.0, inherited from the base model.
• —Recommended agent for using this model: <https://xagentai.net/xagentai-net-coding-agent/>

Citation

@misc{qwen38_bughunter_v7_mlx,
  title  = {Qwen3.8-27B-Uncensored-bughunter-v7 (MLX 4-bit with vision): Apple Silicon build of the bug bounty fine-tune},
  year   = {2026},
  base   = {orcarouter/Qwen3.8-27B-Uncensored},
  method = {LoRA fine-tune (rank=8, lr=1e-5, 1 epoch, QLoRA 4-bit, Unsloth) + mlx_vlm.convert -q},
  data   = {14914 ShareGPT examples distilled from public bug bounty corpus}
}