xGemable-12B-coder-v2-fullft-GGUF

<p align="center"> <img src="xgemable.jpg" alt="xGemable-12B" width="520"/> </p>

xGemable-12B-coder-v2（全參數版・GGUF）

云碩科技（xCloudinfo）以全參數微調（full fine-tune）＋ 全長零截斷配方，在 google/gemma-4-12B-it 底座上蒸餾出的寫程式模型。這是一顆會寫「真的能編譯、真的過測試」程式碼的12B模型。

---

為什麼你該試這顆：一個12B打贏一票對手

所有數字皆為全量164題HumanEval、實際編譯執行判分，非引用宣稱。你可以自己複現。

這顆GGUF（Q4）實測91.5%，仍勝過未訓練底座、大勝同級對手8.6分。 全精度為93.9%； Q4量化損失約2.4分（要更接近全精度可用Q6/Q8）。而且是一顆你自己機器就跑得動的12B。

另在本機ollama（Apple M2 Pro, Q4_K_M）實測一組跨語言基礎題（Python / C# / Rust）：27/27 = 100%， 同樣是實際編譯執行判分。基礎函式題三語言零失誤，難題鑑別度仍以上表HumanEval為準。

---

我們做對了什麼（方法論突破）

多數團隊在強底座上做微調會掉分——我們自己前三次（LoRA低秩 + 截斷）也從90.9% 掉到82.9%， 一度以為「疊加微調必然傷底座」。這顆推翻了這個結論：

• 全參數微調（不是LoRA）——容量夠，學新能力不必擠掉舊能力
• 全長零截斷——完整訓練多輪代理式軌跡，不砍掉解題的長弧
• 執行驗證蒸餾——每一條訓練軌跡都先過真實編譯器與測試，跑不過的一律丟棄

結果：不但沒掉分，還比未訓練底座更強。真兇是「低秩 + 截斷」，不是蒸餾本身。

---

特性

• 多語言：Python為主力，C# / Java / Go / JavaScript / Rust皆可
• 原生function calling / 工具使用：可讀錯誤訊息、多輪迭代
• 地端私有部署：GGUF量化，單機16GB RAM起跑，資料不離內網
• 長上下文1M tokens，開箱即用：YaRN rope-scaling已寫入GGUF metadata（256K原生 × 4）， llama.cpp / LM Studio載入即生效、不需額外參數。整個中型repo、數十份文件一次塞進context做跨檔案推理。 （實際可用長度受你的RAM限制；一般16GB機器建議 -c 32K以內，1M需大記憶體機器。）
• 底座gemma-4-12B-it

---

務必先讀：關思考

底座gemma-4帶「思考通道」。任何推理端都要關思考，否則：

• 直接跑會把token耗在思考、輸出夾雜推敲、分數腰斬；
• 透過 Open WebUI等chat介面會出現「回覆空白／寫不出來」——ollama把思考解析進獨立的 thinking 欄位、content 留空，而UI預設只顯示 content。這不是模型壞了，是沒關思考。

各引擎關思考的對應參數（擇一）：

---

怎麼用

ollama（本機最省事）

# 拉模型
ollama pull hf.co/xCloudinfo/xGemable-12B-coder-v2-fullft-GGUF:Q4_K_M

# 用 chat API（務必帶 think:false，否則 content 會是空的）
curl http://localhost:11434/api/chat -d '{
  "model": "hf.co/xCloudinfo/xGemable-12B-coder-v2-fullft-GGUF:Q4_K_M",
  "messages": [{"role":"user","content":"寫一個 Python 函式 add(a,b)"}],
  "think": false,
  "stream": false
}'

用 Open WebUI：拉好模型後，在對話的Controls／進階參數把 Reasoning／Thinking關掉即可正常輸出。

llama.cpp

# 下載（Q4_K_M 約 7.4GB，16GB RAM 可跑）
hf download xCloudinfo/xGemable-12B-coder-v2-fullft-GGUF \
  xGemable-12B-coder-v2-fullft-Q4_K_M.gguf --local-dir .

# 啟動 server（--chat-template-kwargs 關思考 務必加）
llama-server -m xGemable-12B-coder-v2-fullft-Q4_K_M.gguf -c 8192 --jinja \
  --chat-template-kwargs '{"enable_thinking":false}' \
  --temp 0.2 --top-p 0.95 --top-k 64

啟動後即為OpenAI相容端點（http://localhost:8080/v1），可直接接編輯器外掛或任何OpenAI SDK。

量化版本

建議解碼參數：temperature 0.2、top_p 0.95、top_k 64、repeat_penalty 1.0。

---

能力範圍（誠實揭露）

• 擅長：函式與演演算法實作、單元測試、程式碼解釋與審查、多語言、多輪工具使用。
• 限制：完整倉庫級自動除錯（在整個專案中定位並修復真實bug）尚未達實用門檻——此為當前12B級模型之 共通限制，非本模型獨有。若你的任務是「寫會過測試的函式/模組」，這顆很強；若是「自動修整個repo的bug」， 現階段任何12B都還做不到。

授權

Apache-2.0。 底座 google/gemma-4-12B-it 為Apache-2.0，本模型為其微調衍生版本， 沿用相同授權，可自由商用、修改、再散布。

關於云碩科技

云碩科技（xCloudinfo Corp. Limited）專注於地端私有化人工智慧解決方案。xVITA是云碩自研模型的品牌名稱。

據點：台北、台中、新竹。

其他模型請見 https://hf.co/xCloudinfo。

---

xGemable-12B-coder-v2（全參數版・GGUF） — English

A coding model distilled by xCloudinfo (云碩科技) on the google/gemma-4-12B-it base with a full fine-tune + full-length, zero-truncation recipe. This is a 12B model that writes code that really compiles and really passes tests.

---

Why you should try this one: a 12B that beats a whole field of rivals

All numbers come from the full 164-problem HumanEval, scored by actual compilation and execution; they are not cited claims. You can reproduce them yourself.

This GGUF (Q4) measured 91.5%, still beating the untrained base and beating the same-class rival by a wide margin of 8.6 points. Full precision is 93.9%; Q4 quantization costs about 2.4 points (use Q6/Q8 to get closer to full precision). And it is a 12B that runs on your own machine.

In addition, a set of cross-language basic problems (Python / C# / Rust) was measured on a local ollama (Apple M2 Pro, Q4_K_M): 27/27 = 100%, also scored by actual compilation and execution. There were zero failures on basic function problems across the three languages; for discriminating on hard problems, the HumanEval table above remains the reference.

---

What we got right (methodology breakthrough)

Most teams lose points when fine-tuning on a strong base. Our own first three attempts (low-rank LoRA + truncation) also dropped from 90.9% to 82.9%, and for a while we concluded that "stacked fine-tuning inevitably hurts the base". This model overturns that conclusion:

• Full-parameter fine-tuning (not LoRA): enough capacity, so learning new abilities does not have to squeeze out old ones
• Full length, zero truncation: multi-turn agentic trajectories are trained in full, without cutting off the long arc of problem-solving
• Execution-verified distillation: every training trajectory first passes a real compiler and tests; anything that fails is discarded

The result: not only did it not lose points, it is stronger than the untrained base. The real culprit was "low rank + truncation", not distillation itself.

---

Features

• Multiple languages: Python is the main focus; C# / Java / Go / JavaScript / Rust are all supported
• Native function calling / tool use: can read error messages and iterate over multiple turns
• On-premises private deployment: GGUF quantization, runs on a single machine starting from 16GB RAM, data stays inside the internal network
• Long context of 1M tokens, out of the box: YaRN rope-scaling is written into the GGUF metadata (256K native × 4), so it takes effect on load in llama.cpp / LM Studio with no extra parameters. An entire mid-sized repo or dozens of documents can be put into the context at once for cross-file reasoning. (The length you can actually use is limited by your RAM; on a typical 16GB machine, keeping -c within 32K is recommended; 1M requires a machine with a large amount of memory.)
• Base model gemma-4-12B-it

---

Read first: turn thinking off

The gemma-4 base has a "thinking channel". Turn thinking off on every inference engine; otherwise:

• running it as-is spends tokens on thinking, mixes deliberation into the output, and cuts scores in half;
• through chat interfaces such as Open WebUI, you get "blank replies / it can't write anything": ollama parses the thinking into a separate thinking field and leaves content empty, while the UI shows only content by default. The model is not broken; thinking was not turned off.

The parameter that turns thinking off in each engine (pick one):

---

How to use

ollama (easiest locally)

# 拉模型
ollama pull hf.co/xCloudinfo/xGemable-12B-coder-v2-fullft-GGUF:Q4_K_M

# 用 chat API（務必帶 think:false，否則 content 會是空的）
curl http://localhost:11434/api/chat -d '{
  "model": "hf.co/xCloudinfo/xGemable-12B-coder-v2-fullft-GGUF:Q4_K_M",
  "messages": [{"role":"user","content":"寫一個 Python 函式 add(a,b)"}],
  "think": false,
  "stream": false
}'

With Open WebUI: after pulling the model, turn Reasoning / Thinking off in the chat's Controls / advanced parameters and it will output normally.

llama.cpp

# 下載（Q4_K_M 約 7.4GB，16GB RAM 可跑）
hf download xCloudinfo/xGemable-12B-coder-v2-fullft-GGUF \
  xGemable-12B-coder-v2-fullft-Q4_K_M.gguf --local-dir .

# 啟動 server（--chat-template-kwargs 關思考 務必加）
llama-server -m xGemable-12B-coder-v2-fullft-Q4_K_M.gguf -c 8192 --jinja \
  --chat-template-kwargs '{"enable_thinking":false}' \
  --temp 0.2 --top-p 0.95 --top-k 64

Once started, it is an OpenAI-compatible endpoint (http://localhost:8080/v1) that you can connect directly to an editor plugin or any OpenAI SDK.

Quantized versions

Recommended decoding parameters: temperature 0.2, top_p 0.95, top_k 64, repeat_penalty 1.0.

---

Scope of capabilities (honest disclosure)

• Strong at: function and algorithm implementation, unit tests, code explanation and review, multiple languages, multi-turn tool use.
• Limitations: full repository-level automated debugging (locating and fixing real bugs across an entire project) has not yet reached a practical threshold. This is a limitation shared by current 12B-class models, not one unique to this model. If your task is "write functions/modules that pass tests", this model is strong; if it is "automatically fix the bugs across an entire repo", no 12B can do that at this stage.

License

Apache-2.0. The base google/gemma-4-12B-it is Apache-2.0; this model is a fine-tuned derivative of it and uses the same license, which permits free commercial use, modification, and redistribution.

About xCloudinfo

xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions. xVITA is xCloudinfo's in-house model brand.

Offices: Taipei, Taichung, Hsinchu.

More models: https://hf.co/xCloudinfo