66 Downloads Updated 6 days ago
Agentic coding model for 24 GB GPUs: TeichAI's Gemma-4-31B Fable-5 agent distill (vision + thinking + tools) with a disciplined coding-agent system prompt baked in. Inspect → reproduce → smallest fix → re-verify.
ollama run Barbatos/gemfable-agent:31b
Applications
ollama launch claude --model Barbatos/gemfable-agent:31b
ollama launch codex-app --model Barbatos/gemfable-agent:31b
ollama launch openclaw --model Barbatos/gemfable-agent:31b
ollama launch hermes --model Barbatos/gemfable-agent:31b
ollama launch codex --model Barbatos/gemfable-agent:31b
ollama launch opencode --model Barbatos/gemfable-agent:31b
Models
View all →Readme
gemfable-agent
A ready-to-run agentic coding model built on TeichAI/Gemma-4-31B-Fable-5-Agent-Distill (Apache-2.0) — a Claude Fable 5 agent-trace distill of Gemma-4-31B-it with vision, thinking, and tool calling.
What this repackage adds over the raw GGUF:
- Agentic work method baked into the system prompt (adapted from the Codex-CLI method visible in the CC-BY-4.0 GPT-5.6 Sol coding traces): inspect before editing, reproduce the failure first, evidence-backed root cause, smallest coherent fix, re-run the exact verification. Never alters tests to pass; never claims unverified success.
- Correct Ollama wiring: gemma4 renderer/parser (thinking + tools work), mmproj included (vision works),
repeat_penalty 1(default penalties can send long reasoning into degenerate rambles), sampling per TeichAI’s recommendation (temp 1.0, top_p 0.95, top_k 64), 64K context default.
Benchmarks
Our own runs: Q6_K quantized, llama.cpp on 2×RTX 3090 Ti, greedy (temp 0), pass@1, thinking on, single attempt, seed-42 subsets where marked. Reference is Google’s published Gemma 4 31B (bf16, served). Two handicaps apply to our column: quantization, and subsets. Scripts: github.com/Xza85hrf/gemfable-agent/tree/master/bench-public
| Benchmark | gemfable-agent (Q6_K, local) | Gemma 4 31B (stock, bf16) |
|---|---|---|
| HumanEval / HumanEval+ | 97.6% / 93.3% | — |
| MBPP / MBPP+ | 95.5% / 82.3% | — |
| LiveCodeBench v6 (40) | 85.0% | 80.0% |
| GSM8K (200) | 92.0% | — |
| MATH-500 (100) | 84.0% | — |
| MMLU-Pro (100) | 86.0% | 85.2% |
| GPQA Diamond (198, full) | 80.8% | 84.3% |
| AIME 2026 (30, full) | 83.3% | 89.2% |
| BigBench Extra Hard (100) | 63.0% | 74.4% |
Honest read: a coding-specialized distill. It edges stock on coding (LiveCodeBench +5), matches it on MMLU-Pro, trails 3–6 pts on GPQA/AIME, and takes a real hit on BBEH — the cost of specializing. LiveCodeBench here caps executed tests per problem, so it approximates rather than reproduces the official harness. Not run locally: Codeforces ELO, Tau2, HLE, MRCR-128k, audio (no audio path in this GGUF).
Agentic / applied (our private gates)
| Gate | gemfable-agent | Base Qwen3.6-35B | Qwopus-27B |
|---|---|---|---|
| project-bench (40 hidden acceptance tests) | 40⁄40 | 35⁄40 | 39⁄40 |
| expert fill-in-the-middle (10 executed) | 10⁄10 | 8⁄10 | 9⁄10 |
| agentic build + debug (independently verified) | 2⁄2 | — | — |
It also planned, coded, and self-debugged a complete 1080p animated explainer video (TTS + PIL + ffmpeg pipeline it wrote itself) from just a transcript and a few frames.
Usage
Works out of the box with agent integrations, or as a plain chat/vision/tools model via the OpenAI-compatible API (http://localhost:11434/v1). For the standalone agentic harness (shell + file tools, test-gated finish in one auditable Python file), Modelfile source, and the benchmark scripts:
https://github.com/Xza85hrf/gemfable-agent
~24 GB VRAM at this Q6_K quant. Licenses: base model Apache-2.0; work-method prompt adapted from CC-BY-4.0 traces; harness MIT.