GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
llmfan46
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
llmfan46
Model Tree
Base
yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF
Fine-tuned
this model
Input Modalities
Text
Audio
Image
Video
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Creating these models takes significant time, work and compute. If you find them useful consider supporting me:
| Platform | Link | What you get |
|---|---|---|
| 🎉 Patreon | Monthly support | Priority model requests |
| ☕ Ko-fi | One-time tip | My eternal gratitude |
Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.
| Parameter | Value |
|---|---|
| direction_index | 28.81 |
| attn.o_proj.max_weight | 0.93 |
| attn.o_proj.max_weight_position | 28.53 |
| attn.o_proj.min_weight | 0.87 |
| attn.o_proj.min_weight_distance | 25.58 |
| mlp.down_proj.max_weight | 1.37 |
| mlp.down_proj.max_weight_position | 33.47 |
| mlp.down_proj.min_weight | 0.28 |
| Metric | This model | Original model (gemma-4-12B-coder-fable5-composer2.5-v1-GGUF) |
|---|---|---|
| KL divergence | 0.0467 | 0 (by definition) |
| Refusals | ✅ 9/100 | ❌ 100/100 |
Original:
============================================================
Total questions: 7021
Correct: 5316
Accuracy: 0.7572 (75.72%)
Parse failures: 90
============================================================
Tested subject scores:
Heretic:
============================================================
Total questions: 7021
Correct: 5276
Accuracy: 0.7515 (75.15%)
Parse failures: 97
============================================================
Tested subject scores:
MMLU - Massive Multitask Language Understanding, multiple-choice questions across 57 subjects (math, history, law, medicine, etc.).
GGUF quantizations available here llmfan46/gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-heretic-GGUF.
No matter your GPU. No matter your RAM. If you've got ~4.5 GB of VRAM or unified memory free, you can run your own private, offline coding assistant right now. 🚀 This is the v1 / code edition — distilled from real chain-of-thought so it thinks through a problem before writing the solution. 🧠💻 All local, all yours, no API, no cloud.
A focused fine-tune of Gemma 4 12B on verifiable Python coding data — every training example's reasoning leads to code that actually passed its tests. The result reasons in the open (edge cases, complexity, approach) and then emits a clean, runnable solution. 💚
🚀🔥 IT'S HERE — v2 is OUT NOW! v2 has shipped — the GGUF quants are live and ready to run →
grab v2 here. 🎉
The full safetensors master (build / fine-tune on top) goes up tomorrow. v2 is agentic + coding focused —
the piece v1 was missing.
Here's the result that got me most excited. When I saw v2's tau2-bench telecom result — an agentic tool-use
benchmark where the model has to diagnose → fix → verify, exactly like real terminal/debugging work — I literally got
launched out of my chair (…okay, kidding 😄). The jump in actually solving the problem is wild:
| tau2-bench telecom · local, same harness, Q8_0 | score |
|---|---|
official gemma-4-12B-it (base) | ~15% |
| 🟢 v2 (this release) | ~55% |
The base model tends to give up early (hands the problem off to a human); v2 keeps going and works it the way a much bigger model would. Full benchmark details are in the v2 card now. 🔧
✅ safetensors master (this v1 model) is UP. Full-precision weights are live → yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1 — roll your own GGUF / MLX / AWQ quants or fine-tune straight from the master. 🎉
A community member spotted that this model was reporting only a 131K context window. That turned out to be
the well-known upstream Gemma 4 metadata bug — Google's initial config.json shipped with
max_position_embeddings: 131072 instead of the real 262144 (256K), and that value got baked into a lot of
downstream finetunes and quants (including this one) before it was fixed upstream.
The weights were always fine — it was purely a metadata field. All GGUF quants have been re-patched to the
full 256K context (gemma4.context_length = 262144). Just re-download if you grabbed an earlier copy. 🙏
This is a distillation of two complementary chain-of-thought sources, both over verifiable Python coding tasks (algorithmic / function-level problems that come with deterministic tests):
The recipe: real CoT for the bulk of solid coverage, plus synthetic "second-attempt" CoT to patch the failures — both verified by execution before anything entered training. ✅
| Quant | Size | Vibe |
|---|---|---|
| 🟢 Q2_K | 4.5 GB | tiniest — runs almost anywhere |
| 🟡 Q3_K_M | 5.7 GB | great for 8 GB VRAM — much better than Q2 |
| 🔵 Q4_K_M | 6.87 GB | the sweet spot 👌 (recommended) |
| 🟣 Q6_K | 9.11 GB | near-lossless |
| ⚪ Q8_0 | 11.8 GB | basically full quality |
Rough estimates 🤓 (assumes q8_0 KV cache + ~1.5 GB overhead; use q4_0 KV cache for ≈2× more context!).
Max context is 256K. "—" = won't fit, pick a smaller quant. ✂️
| Your VRAM / unified mem | 🟢 Q2_K (4.5G) | 🟡 Q3_K_M (5.7G) | 🔵 Q4_K_M (6.87G) | 🟣 Q6_K (9.11G) | ⚪ Q8_0 (11.8G) |
|---|---|---|---|---|---|
| 8 GB | ~16K ctx | ~10K | tight (~2–4K) | — | — |
| 12 GB | ~48K | ~38K | ~30K | ~12K | — |
| 16 GB |
💡 Apple Silicon / integrated GPUs with unified memory count too — same numbers, just slower than a dGPU. 💡 Low on room? Drop a quant or switch KV cache to
q4_0and your context roughly doubles.
…-Q4_K_M.gguf) and llama-server from llama.cpp.
⚠️ Needs a recent llama.cpp (this is the
gemma4_unifiedarchitecture — older builds won't load it).
.bat shown — tweak --port, --ctx-size to taste):bat
@echo offcd /d C:\llama.cppllama-server.exe ^-m C:\models\gemma4-coding-Q4_K_M.gguf ^--ctx-size 16384 ^--n-gpu-layers 99 ^--no-mmap ^-fa on ^--cache-type-k q8_0 --cache-type-v q8_0 ^--temp 1.0 --top-p 0.95 --top-k 64 ^--host 0.0.0.0 --port 18080pause
http://localhost:18080 and chat. 🎉 (Tip: bump --ctx-size per the table; use q4_0 KV for more.)Works in LM Studio, Jan, Ollama, etc. — just import the GGUF, pick your quant, go. 🐾
This model thinks in Gemma's native thought channel before answering — exactly how it was trained. Keep
enable_thinking=true (the default chat template handles it). Recommended sampling: temp 1.0, top_p 0.95, top_k 64.
For coding you can also go greedy (temp 0) for more deterministic solutions.
google/gemma-4-12B-it.| mlp.down_proj.min_weight_distance | 26.41 |
| ~80K |
| ~72K |
| ~64K |
| ~44K |
| ~22K |
| 24 GB | ~200K | ~160K | ~128K | ~110K | ~88K |
| 32 GB | 256K (max) 🎉 | 256K | 256K | ~230K | ~190K |