| --- |
| license: apache-2.0 |
| pipeline_tag: image-text-to-text |
| library_name: gguf |
| tags: |
| - code |
| - moe |
| - pxq |
| - pxa-network |
| - multimodal |
| --- |
| |
|  |
|
|
| # PXA Coder 35B |
|
|
| **CODE. BUILD. AUTOMATE.** |
|
|
| > **This model does not run on stock llama.cpp, LM Studio, ollama, or text-generation-webui.** |
| > PXQ4 is a PXA format. You need our engine: |
| > **https://github.com/poisonxa16/pxq_llama.cpp** |
| > If you want something that runs anywhere, this is not that model. If you want 35B on a |
| > 24 GB card, read on. |
| |
| --- |
| |
| ## 1. Meet PXA Coder |
| |
| PXA Coder is built to code, think, and actually get shit done. |
| |
| We're introducing PXA Coder, our next-generation coding model from PXA Network, built for |
| developers who need more than autocomplete and surface-level answers. |
| |
| Under the hood, PXA Coder brings together capabilities from Qwen 3.6 and Qwen 3.8, merged |
| and further trained using the larger "big brother" model as a teacher. We used that |
| foundation to push the model toward stronger coding instincts, better reasoning, and more |
| reliable behavior when tasks get long and complicated. |
| |
| The result isn't meant to be another model that spits out a code snippet and calls it a day. |
| |
| PXA Coder is built for the hard stuff: large codebases, complicated debugging, multi-file |
| changes, architecture decisions, agentic workflows, and problems where the first answer |
| usually isn't the right one. |
| |
| It's built to stay locked onto the goal across long sessions instead of forgetting what it |
| was doing halfway through. |
| |
| Whether you're building an app from scratch, hunting down a nasty bug, refactoring a mess of |
| legacy code, or handing an AI agent an entire project to work through, PXA Coder is built to |
| keep moving. |
| |
| --- |
| |
| ## 2. What PXA Coder Brings |
| |
| ### Built From a Strong Foundation |
| |
| PXA Coder combines the strengths of Qwen 3.6 and Qwen 3.8, then takes that foundation |
| further through additional training guided by the larger model. |
| |
| The idea is simple: take what already works, push it harder, and tune it specifically toward |
| the kind of coding and reasoning workflows we actually care about. |
| |
| ### Built for Real Coding |
| |
| PXA Coder isn't just here to explain code. It's built to write it, modify it, debug it, and |
| build with it. |
| |
| From a single function to an entire application, PXA Coder is designed to work with |
| real-world development workflows. |
| |
| ### Long-Horizon Agentic Work |
| |
| Give it a task that takes 5 minutes — or a task that takes 5 hours. |
| |
| PXA Coder is designed to maintain context, track objectives, work through multiple steps, |
| and recover when things don't go according to plan. |
| |
| ### Handles the Hard Stuff |
| |
| Complex algorithms. Difficult bugs. Weird edge cases. Massive codebases. |
| |
| PXA Coder is designed for problems where "just generate some code" isn't enough. It can |
| break complicated problems down, reason through them, and work toward a solution without |
| constantly losing the plot. |
| |
| ### Multi-File & Codebase-Aware |
| |
| Real projects aren't one Python file. |
| |
| PXA Coder is built to work across files, understand how pieces of a project connect, make |
| coordinated changes, refactor existing code, and chase bugs across an entire codebase. |
| |
| ### Built for Agents |
| |
| PXA Coder is designed to work inside agentic coding environments where the model can inspect |
| files, use tools, run commands, make changes, test its work, and iterate. |
| |
| Think → Build → Test → Fix → Repeat. |
| |
| ### Stays Locked In |
| |
| One of the biggest problems with long AI sessions is drift. |
| |
| The task starts here. |
| |
| Twenty steps later, the model has forgotten why it started. |
| |
| PXA Coder is designed around avoiding that problem — keeping the objective, context, and |
| previous work in view throughout the process. |
| |
| --- |
| |
| ## 3. Performance & Design Philosophy |
| |
| PXA Coder isn't built around giving the flashiest answer in the first five seconds. |
| |
| It's built around getting to the right answer. |
| |
| The model is designed for sustained coding sessions where problems evolve, requirements |
| change, bugs appear, and the first solution needs to be thrown out and rebuilt. |
| |
| Its focus is simple: |
| |
| Better reasoning. |
| Better code. |
| Better execution. |
| |
| The model combines the underlying capabilities of its Qwen-based foundation with additional |
| training designed to make it more useful as a dedicated coding model. |
| |
| The focus is sustained performance: reasoning through difficult problems, navigating larger |
| projects, maintaining context, and continuing to work when the first solution doesn't survive |
| contact with the real codebase. |
| |
| PXA Coder is built to go beyond generating snippets and become an actual coding engine — |
| capable of working through complex projects, making decisions, testing those decisions, and |
| adapting when something breaks. |
| |
| The goal isn't to make an AI that looks like a developer. |
| |
| The goal is to build one that can actually build. |
| |
| --- |
| |
| ## Specifications |
| |
| | | | |
| | ---------------- | ---------------------------------------------------- | |
| | Parameters | 34.7 B total, mixture-of-experts | |
| | Architecture | Qwen MoE, hybrid attention (linear + full, 1-in-4) | |
| | Context | 262,144 trained | |
| | Vocabulary | 248,320 | |
| | Modality | text + vision (ships with an mmproj) | |
| | Quantisation | PXQ4, 19 GB on disk | |
| | Licence | Apache 2.0 | |
| |
| ## Files |
| |
| | File | Size | What it is | |
| | ---- | ---- | ---------- | |
| | `PXA-Coder-35B-v2-PXQ4.gguf` | 19 GB | the model | |
| | `mmproj-PXA-Coder-35B-v2-f16.gguf` | 899 MB | vision projector — only needed for image input | |
| |
| ## Running it |
| |
| ```bash |
| git clone https://github.com/poisonxa16/pxq_llama.cpp |
| cd pxq_llama.cpp |
| cmake -B build -S . -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="70;80;86;89;90" |
| cmake --build build -j --target llama-server |
| |
| ./build/bin/llama-server \ |
| -m PXA-Coder-35B-v2-PXQ4.gguf \ |
| --mmproj mmproj-PXA-Coder-35B-v2-f16.gguf \ |
| -ngl 99 -c 32768 |
| ``` |
| |
| Drop `--mmproj` if you only want text. Lower `-c` if you are tight on VRAM — the KV cache, |
| not the weights, is what will run you out. |
| |
| **Building on a machine with no GPU present?** The CUDA toolkit's driver stub has to be on |
| the link line or you will get `undefined reference to cuMemCreate` and friends. Paths spelled |
| out in full so nothing depends on your shell: |
|
|
| ```bash |
| sudo ln -sf /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1 |
| cmake -B build -S . -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=60 -DLLAMA_CURL=OFF \ |
| -DCMAKE_EXE_LINKER_FLAGS="-L/usr/local/cuda/lib64/stubs -Wl,-rpath-link,/usr/local/cuda/lib64/stubs" \ |
| -DCMAKE_SHARED_LINKER_FLAGS="-L/usr/local/cuda/lib64/stubs -Wl,-rpath-link,/usr/local/cuda/lib64/stubs" |
| cmake --build build -j --target llama-server llama-cli |
| ``` |
|
|
| ## Why PXQ4 |
|
|
| PXQ4 stores weights as 64-row panels with a per-row fp16 anchor and a shared sub-block |
| scale, rather than per-block scales the way K-quants do. On this model that is 19 GB against |
| 69 GB at f16 — 35B fits on a single 24 GB card with room for context. |
|
|
| The format is the reason for the engine requirement. It is not a repackaged GGUF quant; the |
| dequant and matmul kernels are ours. |
|
|
| ## Safety |
|
|
| **This model has had its refusal behaviour deliberately reduced.** It will attempt requests |
| that a stock instruct model declines. |
|
|
| That is the intended design — it exists to be a working tool that does not refuse benign |
| technical work because a variable is named `kill_process`. It is not a safety-aligned |
| assistant, it has no guardrails you can rely on, and you are responsible for what you do |
| with it and for whatever you put in front of users. |
|
|
| ## Evaluation status |
|
|
| Being straight about this, because the model is new: |
|
|
| - **Coding, reasoning, instruction following, vision** — exercised and working. |
| - **Long-context retention is measured, and it holds.** Needle-in-a-haystack against the |
| served model, with the fact placed both early and late in the window: |
|
|
| | Prompt tokens | Early recall | Late recall | |
| | ------------- | ------------ | ----------- | |
| | 10,893 | hit | hit | |
| | 43,578 | hit | hit | |
| | 163,431 | hit | hit | |
| | 250,074 | hit | hit | |
|
|
| 100% at every depth tested, up to 250k tokens. |
| - **Agentic and multi-file workflows are a design target, not yet a benchmarked result.** |
| If you put it through a real agent loop, we want the numbers — including bad ones. |
|
|
| ## Credits |
|
|
| Qwen-based foundation, Apache 2.0. Quantisation format, engine, and post-training by |
| PXA Network. Licensed Apache 2.0; this is a modified work. |
|
|
| --- |
|
|
| **PXA Coder** — Built by PXA Network for developers, builders, engineers, and anyone who'd |
| rather ship than stare at a blank terminal. |
|
|
| *Your ideas. Our intelligence. Real results.* |
|
|
