DEV.co
Open-Source LLM · Jackrong

Qwopus3.6-27B-Coder-MTP-GGUF

Qwopus3.6-27B-Coder is a 27-billion-parameter open-source language model fine-tuned for coding and agent tasks. It builds on Qwopus3.6-27B-v2 (itself a reasoning-enhanced version of Qwen3.6-27B) and adds specialized training for code generation, tool calling, debugging, and multi-turn agentic workflows. The model reports 67.0% on SWE-bench Verified (500 tasks, thinking-off mode). It is distributed as GGUF (quantized format) under Apache 2.0, is ungated, and has ~304k downloads. The developer warns this is an experimental community release not yet fully safety-evaluated across broad domains.

Source: HuggingFace — huggingface.co/Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
304.1k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperJackrong
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads304.1k
Likes324
Last updated2026-06-24
SourceJackrong/Qwopus3.6-27B-Coder-MTP-GGUF

What Qwopus3.6-27B-Coder-MTP-GGUF is

Dense 27B transformer trained via Trace Inversion (reasoning trajectory reconstruction from Claude Opus data), LoRA/SFT refinement, and agent trajectory learning. Base is Qwen3.6-27B (Alibaba); specialization layers target agentic code generation, structured tool calling, and repository-level reasoning. GGUF quantization enables single-GPU deployment. Achieved 87.43% MMLU-Pro and 75.25% SWE-bench on v2 base; Coder variant reports 67.0% SWE-bench Verified (500 tasks, no-thinking mode, Q5_K_M quantization). Context length, exact parameter count, and tokenizer details not explicitly stated. Training data includes Claude-opus synthetic reasoning traces and agent-reasoning datasets.

Quickstart

Run Qwopus3.6-27B-Coder-MTP-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Jackrong/Qwopus3.6-27B-Coder-MTP-GGUF")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.

Deployment

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

Autonomous code agents & SWE automation

SWE-bench Verified 67% performance and explicit agentic training make this suitable for building autonomous or semi-autonomous software engineering agents that must call tools, iterate, and handle multi-step repository tasks.

Self-hosted code completion & debugging

27B dense model with GGUF quantization supports single-GPU or CPU deployment. Useful for organizations needing on-premise code assistance without cloud dependency or third-party API exposure.

Research & evaluation of open agentic LLMs

Trace Inversion and published benchmark results (SWE-bench 67%) allow researchers to study open-source reasoning-enhanced models, compare against proprietary Claude/OpenAI agents, or prototype novel agent frameworks.

Running & fine-tuning it

27B dense model. ESTIMATE: Q5_K_M quantization ~18–22 GB VRAM (based on typical 5-bit GGUF overhead). Full precision (bfp16) would require ~54 GB. Single A100 (80GB), dual RTX 6000 (48GB ea.), or CPU with 64+ GB RAM + appropriate quantization. Inference speed on GPU ~5–15 tokens/sec (unverified; depends on hardware, quantization, and batch size). Exact tokenizer context length not stated in data provided.

Model card mentions LoRA and SFT in tags and training stack. GGUF is a quantized inference format; full fine-tuning would require loading/converting to native weights (bfp16/fp32). LoRA adapters for domain-specific code or tool-use patterns are plausible if base weights (not GGUF artifact) are available. Unsloth is referenced in tags; this suggests fast fine-tuning harnesses may be available. Requires evaluation of whether full weights are published separately from GGUF.

When to avoid it — and what to weigh

  • Guaranteed production safety & compliance — Model card explicitly states 'not undergone full safety evaluation.' Avoid for regulated industries (healthcare, finance, legal) without additional safety vetting and domain-specific testing.
  • Low-latency real-time inference at scale — 27B model size and GGUF quantization imply higher latency than smaller 7B models or pruned variants. Requires careful hardware planning if sub-100ms response times are critical.
  • Guaranteed multilingual coding parity — Tags list 6 languages (en, zh, es, ru, ja) but SWE-bench (English) is the only completed benchmark. Code performance in non-English codebases or mixed-language repos remains unknown.
  • Commercial deployment without legal review — Apache 2.0 is permissive, but model is experimental/community-released and unaudited for commercial liability or data leakage. Requires legal/compliance sign-off before production use in revenue-critical systems.

License & commercial use

Apache 2.0 (https://www.apache.org/licenses/LICENSE-2.0). Clear OSI-approved permissive license. Permits commercial use, modification, distribution, and private use with attribution and no-warranty disclaimer. No restrictions on model output or commercial applications stated in license text.

Apache 2.0 permits commercial deployment without license restrictions. However: (1) model is explicitly experimental/community-release and not safety-audited, creating liability risk if harm or data leakage occurs; (2) base model (Qwen3.6-27B) is from Alibaba (verify Alibaba's own commercial terms apply transitively); (3) training data includes Anthropic's Claude-opus synthetic traces and other datasets—verify downstream licensing of training data is compatible with your commercial intent. Recommend legal/compliance review before revenue-critical deployment.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceModerate
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM risks apply: (1) model may generate plausible but incorrect code or hallucinated package imports; (2) agentic tool-calling variant increases risk of unintended function execution if tool definitions/constraints are misconfigured; (3) training data includes synthetic Claude-Opus traces and agent trajectories—potential for subtle prompt-injection or jailbreak patterns to be present; (4) no mention of red-teaming, adversarial testing, or safety filtering. On-premise deployment avoids cloud-provider data exposure but places security responsibility on operator. Always sandbox tool execution and validate model outputs before deployment.

Alternatives to consider

DeepSeek-Coder-33B (or DeepSeek-V3)

Larger open coder model with stronger SWE benchmarks; better long-context (4k–32k); similar Apache/MIT licensing. Trade-off: higher compute cost, less reasoning-trajectory transparency.

Qwen2.5-Coder (32B, 7B variants)

Official successor from Alibaba; likely more stable, broader benchmarking, native long-context support. Trade-off: newer, may have less community adoption; Qwopus is experimental/community-optimized.

CodeLlama-70B or Llama 3 Instruct (70B)

Proven production code models with broad deployment experience and strong general-instruction-following. Trade-off: not specialized for agentic tool-use; larger compute footprint; less reasoning-trajectory augmentation.

Software development agency

Ship Qwopus3.6-27B-Coder-MTP-GGUF with senior software developers

Qwopus3.6-27B-Coder is a strong candidate for research, on-premise code assistance, and agent automation. Pair it with llama.cpp or Ollama for quick local deployment, or use our engineering services to integrate it into your SWE toolchain. Consult legal/compliance before production use.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Qwopus3.6-27B-Coder-MTP-GGUF FAQ

Can I use this commercially without additional licensing?
Apache 2.0 permits commercial use. However, the model is labeled experimental/community-release and unaudited for safety/liability. Base model is Alibaba's Qwen—verify no Alibaba commercial restrictions apply. Training data includes Anthropic Claude traces—confirm licensing compatibility. Strongly recommend legal/compliance review before revenue-critical deployment.
What hardware do I need to run this?
ESTIMATE: Q5_K_M quantization requires ~18–22 GB VRAM (A100 80GB, dual 48GB cards, or high-end CPU). Full precision (bfp16) requires ~54 GB. Use llama.cpp, Ollama, or vLLM for serving. Exact latency/throughput depends on hardware and quantization; not published.
What is the main difference between this and Qwopus3.6-27B-v2?
v2 is general-purpose reasoning-enhanced; Coder variant adds specialized fine-tuning for agentic code generation, tool-calling, and SWE tasks. SWE-bench Verified: v2 achieved 75.25%, Coder achieves 67.0% (no-thinking mode). Trade-off: Coder is more specialized but slightly lower absolute performance on pure SWE; stronger on agent orchestration and tool use.
How mature is this for production use?
Not recommended for production without mitigation. Model card states 'experimental community release' with no full safety evaluation. Suitable for research, evaluation, and internal tooling with proper sandboxing. For customer-facing or regulated use, add safety vetting, test coverage, and fallback mechanisms.

Custom software development services

DEV.co helps companies turn open-source tools like Qwopus3.6-27B-Coder-MTP-GGUF into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Ready to Deploy an Open Agentic Code Model?

Qwopus3.6-27B-Coder is a strong candidate for research, on-premise code assistance, and agent automation. Pair it with llama.cpp or Ollama for quick local deployment, or use our engineering services to integrate it into your SWE toolchain. Consult legal/compliance before production use.