DEV.co
Open-Source LLM · yuxinlu1

gemma-4-12B-coder-fable5-composer2.5-v1-GGUF

Gemma-4-12B-Coder is a fine-tuned 12B parameter coding model optimized for local inference. It combines real chain-of-thought reasoning from passing test cases with synthetic reasoning traces, and runs on 4.5–12 GB VRAM depending on quantization. The model is distributed as GGUF (quantized) weights for llama.cpp and similar runtimes. Licensed under Apache 2.0, it is intended for Python/algorithmic coding tasks and has reduced refusals compared to the base model.

Source: HuggingFace — huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
675k
Downloads (30d)

Key facts

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

FieldValue
Developeryuxinlu1
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads675k
Likes2.6k
Last updated2026-06-19
Sourceyuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF

What gemma-4-12B-coder-fable5-composer2.5-v1-GGUF is

A distilled derivative of google/gemma-4-12B-it, fine-tuned on dual chain-of-thought data: (1) Composer 2.5 real CoT verified by execution, (2) Fable 5 synthetic CoT for failure-case recovery. Uses Gemma 4's native thinking channel. Context window is 256K (recently corrected upstream metadata bug). Distributed in GGUF quantizations (Q2_K through Q8_0) for local inference. No gating. Requires recent llama.cpp (gemma4_unified architecture support).

Quickstart

Run gemma-4-12B-coder-fable5-composer2.5-v1-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="yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-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

Local coding assistant for individual developers

Run a private, offline coding helper on a laptop or desktop with 8–16 GB RAM. Q4_K_M (6.87 GB) quant recommended for ~30K context on 16 GB systems. No cloud API, no data leaving your machine.

Small team / edge deployments

Deploy in resource-constrained environments (e.g. on-prem offices, remote sites, or embedded applications). Q2_K (4.5 GB) or Q3_K_M (5.7 GB) fits most 8–12 GB systems. Reduced latency vs. cloud; lower operational cost.

Fine-tuning base for domain-specific coding tasks

Full-precision safetensors weights are published separately. Use as a teacher model or starting point for LoRA/QLoRA adaptation to specialized coding domains (e.g., SQL, Rust, domain-specific languages). Reasoning traces in training data may transfer.

Running & fine-tuning it

Minimum ~4.5 GB VRAM/unified memory (Q2_K quant, ~2–4K context window on 8 GB systems). Recommended: 12–16 GB for Q4_K_M with ~30–64K context. Context window is 256K max, but sustained high-context inference requires 24+ GB (Q4_K_M) or 32 GB (Q8_0 full quality). ESTIMATE: Verify exact context capacity in your target environment; card provides rough guidance but actual throughput depends on KV cache quantization, CPU/GPU architecture, and inference library settings.

Full-precision safetensors weights published separately (yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1). LoRA/QLoRA adaptation is feasible; card does not explicitly document LoRA compatibility but Gemma 4 derivatives typically support it. Start from safetensors, not GGUF. Train on verifiable (test-gated) data to preserve reasoning quality. Card notes training data is task-focused; domain-specific fine-tuning likely requires task-specific annotation.

When to avoid it — and what to weigh

  • Production safety-critical systems without guardrails — Model explicitly lacks safety alignment ('not safety-aligned'). Training data is task-focused with no refusal patterns. Do not deploy to user-facing production without additional safety filtering, input validation, and output vetting.
  • Multi-language or general-knowledge tasks — Specialized for Python/algorithmic coding. Factual recall and non-English capability are unknown and likely weak. Card notes 'general-knowledge facts/numbers should still be double-checked.' Use a broader-trained model for QA, translation, or general assistance.
  • Real-time, ultra-low-latency requirements — GGUF inference on CPU or modest GPU is slower than cloud APIs or dedicated inference endpoints. First-token latency and throughput depend heavily on hardware; not suitable for sub-100ms SLA requirements.
  • Scenarios requiring proof of provenance or auditability — Personal/hobby project ('shared as-is, no warranty'). Limited transparency on training data curation (Composer 2.5, Fable 5 datasets not independently verified). No published benchmarks for non-tau2 domains. Difficult to audit or reproduce.

License & commercial use

Apache License 2.0. Google released Gemma 4 (unlike Gemma 1/2/3) under Apache 2.0, and this fine-tune inherits that license. License text: https://ai.google.dev/gemma/apache_2. Full permissions to use, modify, redistribute, and prepare derivative works, provided Apache 2.0 terms are retained.

Apache 2.0 is a permissive OSI license that explicitly permits commercial use, modification, and redistribution, provided you retain copyright notices and license text. No restrictions on proprietary derivatives or commercial applications. However, verify Google's Gemma 4 commercial policy and any downstream dependencies (e.g., training data sources). This is a personal/hobby project ('shared as-is, no warranty'); assess your organization's risk tolerance for using unsupported derivatives in production.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

No inherent vulnerabilities disclosed. Model lacks safety alignment; reduced refusals mean it may produce harmful, biased, or nonsensical code without guardrails. GGUF inference is local/offline (reduces network attack surface) but client-side input/output validation is the user's responsibility. Verify code generated by the model before execution (especially in multi-user or untrusted environments). Training data provenance (Composer 2.5, Fable 5) is opaque; no third-party security audit published.

Alternatives to consider

Llama 2 / Llama 3.2 (7B/8B) + custom coding fine-tune

Open-source, well-documented, broader community support. Requires your own fine-tuning but provides full control over training data and safety alignment. Slightly smaller default models; fewer thinking traces.

DeepSeek-Coder-6.7B / CodeLlama-7B

Purpose-built coding models with stronger non-Python language support and published benchmarks. More mature community fine-tuning examples. Trade-off: less chain-of-thought reasoning exposed.

Proprietary APIs (OpenAI Codex, Claude, Vertex AI Code Assist)

Broader capability, enterprise safety alignment, managed scaling. Higher latency/cost and data residency concerns; no offline option. Use if local deployment is infeasible or multi-language support is critical.

Software development agency

Ship gemma-4-12B-coder-fable5-composer2.5-v1-GGUF with senior software developers

Download a GGUF quant (Q4_K_M recommended), spin up llama-server, and start generating verifiable Python solutions offline. No API key, no cloud latency, full control. Review the full model card for context window estimates and sampling guidance.

Talk to DEV.co

Related on DEV.co

Explore the category and the services that help you build with it.

gemma-4-12B-coder-fable5-composer2.5-v1-GGUF FAQ

Can I use this model in a commercial product without paying royalties?
Yes, Apache 2.0 permits unrestricted commercial use. However, (1) verify Google's Gemma 4 commercial terms independently, (2) this is a personal hobby project ('shared as-is, no warranty'), and (3) you must include the license text and copyright notices in your distribution. Assess your organization's risk tolerance for unsupported third-party derivatives.
I have 8 GB of VRAM. What context window can I expect?
With Q2_K quant (~4.5 GB), card estimates ~16K context on 8 GB with default KV cache. Using q4_0 KV cache (lower precision) instead of q8_0 roughly doubles context to ~32K. For Q3_K_M, expect ~10K default. Trade-off: lower KV precision may affect reasoning quality in long contexts.
Do I need to use 'thinking mode'?
Model uses Gemma 4's native thinking channel and was trained with it enabled. Recommended: keep `enable_thinking=true` (default in chat templates) for coding tasks; thinking traces improve reasoning quality. For greedy/deterministic output, you can set `temp=0`, but intermediate reasoning steps are still computed.
Is this safe to deploy to users without additional safeguards?
No. Model explicitly lacks safety alignment and has reduced refusals. Add input validation, output filtering, rate-limiting, and content moderation before production deployment. Test on adversarial inputs and code injection attempts.

Custom software development services

Adopting gemma-4-12B-coder-fable5-composer2.5-v1-GGUF is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy a Local Coding Assistant?

Download a GGUF quant (Q4_K_M recommended), spin up llama-server, and start generating verifiable Python solutions offline. No API key, no cloud latency, full control. Review the full model card for context window estimates and sampling guidance.