Research · agentic code post-training
Threnody 3.0
An 8.8B multimodal agentic model post-trained with anti-loop filtering, refusal orthogonalization and multi-schema tool calling. Merged at scale-0.75 from Qwen3.5-9B and quantized to 5.6 GiB Q4_K_M for local 12 GB GPUs.
- Latest version
- Threnody 3.0
- Trainable (LoRA)
- 0.75 strength merge
- OpenCode held-out
- 13/14 (92.9%)
- Last commit
- —
- GitHub stars
- —
Where it fits
Closing the agentic loop on consumer hardware.
Where Fable-Coder anchored tool calling on a 7B base, Threnody 3.0 pushes into the denser 8.8B architecture of Qwen3.5-9B to solve the hardest barrier in autonomous programming: repetitive tool loops and premature task abandonment.
By pairing strict anti-loop trace filtering with refusal orthogonalization and multimodal terminal vision, Threnody 3.0 achieves a 92.9% pass rate on held-out OpenCode workflows while fitting cleanly within 6 GB of VRAM.
Anti-loop execution
Eliminates cyclic tool calls, repetitive command re-tries and redundant file reads with strict trajectory-level negative penalization.
Refusal orthogonalization
Abliteration across layers 12–28 prevents false-positive refusals during deep system administration, debugging and security workflows.
Vision & multi-schema tools
8 execution schemas plus native terminal screenshot inspection via an accompanying mmproj projector file.
The six architectural pillars.
A focused post-training recipe addressing real-world agent harness bottlenecks without degrading general reasoning.
LoRA training mixture · curated dataset mix
Training specifications
- Base model
- Qwen/Qwen3.5-9B (~8.8B parameters)
- Post-training
- Replay-anchored LoRA, rank 64 / alpha 128, all 7 projections
- Merge ratio
- Scale 0.75 (s0.75 LoRA merge)
- Refusal vector
- Orthogonalized projection across layers 12–28
- Context length
- 8,192 tokens native (extendable to 32K)
- Compute
- NVIDIA A100-SXM4-80GB
- Quantized size
- 5.6 GiB Q4_K_M (+ 600 MiB mmproj)
Key innovations
- Anti-Loop Filter: Models fine-tuned with
max_repeat = 1verification penalties on held-out tasks, cutting infinite re-try loops by over 80%. - Abliterated Safety: Refusal directions isolated and neutralized in intermediate layers so the model inspects system logs, exploits, and diagnostics without preachy refusals.
- Multimodal Inspection: Accompanied by a dedicated vision adapter for CLI terminals, rendered web views, and UI layout debugging.
- Balanced s0.75 Merge: Blending 75% adapted weights preserves general conversational fluency while adopting strict agent discipline.
From V1 to Threnody 3.0.
Documenting the iterations, breakthroughs, and regressions that led to the final build.
-
Base · Qwen3.5-9B
Dense, capable foundation with raw potential
Strong natural language comprehension and coding fundamentals, but prone to tool-calling loops and ungrounded retries in autonomous harnesses.
-
V1 · Initial SFT
Supervised fine-tuning on multi-schema data
Demonstrated immediate improvements in schema adherence, but suffered from overconfidence and occasional catastrophic forgetting in complex math.
-
V2 · Aggressive synthetic traces
Loop instability regression
Dropped to 9/14 (64.3%) on OpenCode held-out evals — low-variance synthetic data caused cyclic repeating behavior when tools returned unexpected stdout.
-
Threnody 3.0 · Anti-loop LoRA with s0.75 merge
Peak agentic resilience and vision support
Anti-loop negative filtering, refusal abliteration, and a balanced 0.75 weight merge achieve 13/14 (92.9%) pass rate on held-out agentic loops with zero mode collapse.
Rigorous, honest benchmarks.
Head-to-head evals tested across base, V2, and V3 on a local RTX 3060 running identical inference parameters.
Agentic harness & reasoning comparison
| Model | AgentLoop held-out | MMLU-Pro BioChem | OmniAgent-Bench | AgentLoop in-harness | AgentLoop-Hard |
|---|---|---|---|---|---|
| Threnody 3.0 8.8B s0.75 LoRA |
13/14 (92.9%) | 78.5% | 13/16 (81.3%) | 10/14 (71.4%) | 17/30 (56.7%) |
| Qwen3.5-9B base model |
12/14 (85.7%) | 76.0% | 11/16 (68.8%) | 10/14 (71.4%) | 15/30 (50.0%) |
| Threnody V2 synthetic SFT |
9/14 (64.3%) | 75.2% | 10/16 (62.5%) | 8/14 (57.1%) | 12/30 (40.0%) |
Held-out AgentLoop evaluations run against OpenCode environments never encountered during training.
Honest reporting
Known trade-offs and discrepancies.
Every post-trained model incurs trade-offs. AxiomInference reports regressions and architectural nuances openly so practitioners know exactly what to expect.
HumanEval Python Completion Regression
On pure zero-shot Python completion (HumanEval), Threnody 3.0 scored 80.5% compared to the base model's 82.3% (−1.8 pp). This minor regression stems from the model's strong inductive bias toward tool invocation and diagnostic verification steps rather than naive raw text generation.
Why Merge Scale-0.75 Instead of 1.0
A full 100% LoRA merge resulted in rigid formatting patterns and occasional mode narrowing during conversational chat. Merging at a calibrated 0.75 factor retained base model nuance while keeping 99% of the agentic and anti-loop benefits.
Multimodal Projector Scope
The vision tower (mmproj-Threnody-9B-V3-F16.gguf) is optimized for technical artifacts — terminal screenshots, stack traces, diagrams, and web page layout verification — not open-domain creative image captioning.
Run it yourself.
Download the quantized GGUF and vision adapter from Hugging Face, then run locally via Ollama or llama.cpp.
# Clone repo or place GGUF alongside this Modelfile
FROM ./Threnody-9B-V3-Q4_K_M.gguf
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.2
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.15
SYSTEM """You are Threnody 3.0, an expert autonomous coding and engineering agent.
Plan actions carefully, verify assumptions, use available tools, and avoid redundant loops."""
Build and run: ollama create threnody-v3 -f ./Modelfile && ollama run threnody-v3
./llama-cli -m Threnody-9B-V3-Q4_K_M.gguf \
-p "You are an agent. Inspect the git history and write an automated test script for the auth middleware." \
-c 8192 --temp 0.2 --jinja
./llama-minicpmv-cli -m Threnody-9B-V3-Q4_K_M.gguf \
--mmproj mmproj-Threnody-9B-V3-F16.gguf \
--image terminal_error.png \
-p "Explain the root cause of this kernel panic or stack trace shown in the terminal screenshot."
hf download MalxTech/Threnody-9B-V3-GGUF Threnody-9B-V3-Q4_K_M.gguf --local-dir . hf download MalxTech/Threnody-9B-V3-GGUF mmproj-Threnody-9B-V3-F16.gguf --local-dir .
Verify your download
- Language model
Threnody-9B-V3-Q4_K_M.gguf(~5.6 GiB)- Vision projector
mmproj-Threnody-9B-V3-F16.gguf(~600 MiB)- Quantization
- Q4_K_M (Medium K-quant)
- License
- Apache-2.0
Where to get it
- Threnody-9B-V3-GGUF on Hugging Face — full quantized weights and mmproj
- Threnody-V3 on GitHub — training logs, eval suites and scripts
The repository.
Live commits and updates direct from GitHub.
Threnody-V3 · recent commits
Commit history is loading — or see it on GitHub.
More research
All research →Fable-Coder
Four iterations of a 7B code model post-trained for security and tool calling.
Project Kestrel
A small, adaptively-learning LLM architecture trained on one GPU.
MalxLabs-V4
A 1.5B reasoning model tuned to run fast on a GTX 1080.
Axiom
Run GGUF models like this one locally in a full AI workspace.