[
    {
        "role": "user",
        "content": "does%20framework%20desktop%20ai%20pc%20come%20pre-loaded%20with%20linux%3F"
    },
    {
        "role": "assistant",
        "content": "No \u2014 the Framework Desktop PC does *not* come pre-loaded with Linux by default. Here\u2019s what I found from recent and reliable sources:\n\n---\n\n### What does it ship with?\n\nAccording to Tom\u2019s Hardware, the Framework Desktop comes standard with **Windows 11 Home** in its review unit. ([tomshardware.com](https://www.tomshardware.com/desktops/gaming-pcs/framework-desktop-review?utm_source=openai))\n\n---\n\n### Is Linux installable / supported?\n\nYes \u2014 you *can* use Linux on the Framework Desktop. Framework explicitly states in their product listings that you can \u201cBring your preferred OS, including Linux.\u201d ([frame.work](https://frame.work/desktop?utm_source=openai))\n\nThey also outline compatibility on the Framework site for major Linux distributions like Ubuntu and Fedora, listing kernel version requirements (minimum 6.11, recommended 6.15+, best experience on the latest kernel) and confirming that Wi-Fi and Bluetooth work out of the box. ([frame.work](https://frame.work/desktop?tab=linux&utm_source=openai))\n\n---\n\nIf you want, I can also check whether there\u2019s a version you can order *with* Linux pre-installed (if that\u2019s something you\u2019d prefer)."
    },
    {
        "role": "user",
        "content": "how%20well%20does%20debian%2013%20work%20on%20framework%20desktop%20pc%3F"
    },
    {
        "role": "assistant",
        "content": "Debian 13 (code-name *Trixie*) can work well on Framework hardware, but there are a few caveats\u2014some components work perfectly out of the box, others require tweaking or installing non-free firmware. Below is a detailed breakdown of how well it performs, what issues to expect, and how to address them.\n\n---\n\n## \u2705 What works well\n\nFrom the official **Debian Wiki \u201cInstalling Debian On Framework\u201d** pages:\n\n- For **Framework Laptop 13 with AMD Ryzen 7040-series** mainboards, Debian 13 supports most core hardware features by default\u2014booting with the standard kernel, drive detection, sleep/suspend, hibernation, etc. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))  \n- Components like the wireless (WiFi), keyboard hotkeys, touchpad, and ambient light sensor all show up as working. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))  \n- Audio works, though there\u2019s a known buzzing issue in some headphone or speaker setups if the audio chip\u2019s power-saving settings leave output pins floating. That can be mitigated by disabling power saving for the `snd_hda_intel` module. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))\n\n---\n\n## \u26a0\ufe0f Known issues and limitations\n\nSome things don\u2019t work perfectly or require extra steps:\n\n- The **AMD integrated GPU** (RDNA3 architecture) sometimes needs *non-free firmware* for full functionality. Debian 13 *does* include it by default, but issues can still happen if your firmware is out of date. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))  \n- Features like **OpenGL/Vulkan support** are noted as \u201cnot OK\u201d on certain models as of the Debian Wiki status page. For GPU-accelerated graphics or playing videos with hardware decoding, you may hit performance or stability issues. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))  \n- BIOS/UEFI version matters. Early versions (for example, 3.03) can cause issues; updating to something more recent (e.g. version 3.17 or later for some AMD Framework 13 models) is recommended. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))  \n- On laptops with newer designs or features (for example, tablet-mode or unusual sensors), Debian 13's relatively conservative kernel (initial release was 6.12) means some newer hardware quirks might not be fully supported yet. Users report missing drivers or sensors not behaving as expected. ([reddit.com](https://www.reddit.com/r/framework/comments/1p135cn/framework_12_tablet_mode_is_not_working_on_debian/?utm_source=openai))\n\n---\n\n## \ud83d\udd27 What to do to get the best experience\n\nIf you're planning to run Debian 13 on a Framework device, here are steps and tips that help:\n\n1. **Ensure `non-free` firmware is enabled**  \n   Many needed firmware blobs\u2014especially for wireless, graphics, audio output\u2014are in the *non-free* Debian repositories. These are installed by default with Debian 13, but double-check if you disabled non-free in setup. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))\n\n2. **Update BIOS / firmware**  \n   As noted above, updating BIOS/UEFI (and firmware via `fwupd`) can resolve many issues, especially with battery behavior, thermal control, display performance. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))\n\n3. **Use newer kernels or enable backports**  \n   If you need better GPU/multimedia support, the newer kernels from backports (for example for Debian 12, but presumably also possible for 13 backports) help. For bleeding-edge features, unofficial or newer kernel images may be beneficial. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))\n\n4. **Check community guides**  \n   The Debian Wiki pages for *Framework/Laptop13/AMD-7040 Series* etc. are very helpful. They include workarounds, example config files, and tips for audio, display, power management, etc. ([wiki.debian.org](https://wiki.debian.org/InstallingDebianOn/FrameWork/Laptop13/AMD_7040_Series?utm_source=openai))\n\n---\n\n## \ud83e\uddd0 Verdict\n\nIf your Framework desktop/laptop is based on mainstream Intel or AMD hardware (especially AMD 7040 or similar), Debian 13 will be solid for everyday use\u2014coding, web, office, light media. Expect a few speed bumps if you push GPU acceleration, gaming, or very new hardware features.\n\nIf you want the most cutting-edge hardware support and stability, you might consider a distro with newer kernels (e.g. Fedora, Ubuntu, or a Debian derivative with rolling components), or use Debian but lean into backports and firmware updates.\n\n---\n\nIf you tell me exactly which Framework model you have (CPU/mainboard, version, etc.), I can give a more specific sense of what works and what might still need tweaks."
    },
    {
        "role": "user",
        "content": "what%20is%20the%20name%20of%20the%20framework%20pc%20designed%20for%20ai%3F"
    },
    {
        "role": "assistant",
        "content": "The Framework PC designed with a focus on AI workloads is called the **Framework Desktop**. ([techcrunch.com](https://techcrunch.com/2025/02/26/frameworks-first-desktop-pc-is-ready-for-gaming-and-local-ai-inference/?utm_source=openai))\n\nHere are some key details:\n\n- It uses **AMD\u2019s Ryzen AI Max processors**, also known as the *Strix Halo* APU family. ([techcrunch.com](https://techcrunch.com/2025/02/26/frameworks-first-desktop-pc-is-ready-for-gaming-and-local-ai-inference/?utm_source=openai))  \n- There are two CPU options: the **Max 385** (entry-level) and the **Max+ 395** (high-end). ([framework.wiki](https://framework.wiki/docs/computers/desktop/amd-ryzen-ai-max/?utm_source=openai))  \n- RAM options go up to 128 GB of LPDDR5x, though the memory is soldered in and not user-upgradable. ([framework.wiki](https://framework.wiki/docs/computers/desktop/amd-ryzen-ai-max/?utm_source=openai))  \n\nIf you meant a different model (like one by DeepComputing or MetaComputing under \u201cFramework compatibility\u201d), I can also pull up that info."
    },
    {
        "role": "user",
        "content": "so%20if%20it%20is%20using%20amd%20gpu%2C%20then%20no%20cuda%2C%20seems%20like%20ai%20would%20be%20limited%3F%20do%20most%20models%20run%20just%20as%20well%3F"
    },
    {
        "role": "assistant",
        "content": "You're right that CUDA is NVIDIA\u2019s proprietary GPU compute platform and has long been dominant for training and using many AI models. But AMD has made big strides with its ROCm stack, and many modern AI models *do* run well on AMD GPUs\u2014including newer ones with RDNA3/4 or AI/APU architectures. The trade-offs tend to be in performance fine-tuning and ecosystem maturity rather than outright impossibility. Here's a breakdown of how it all shapes up:\n\n---\n\n## \u2699\ufe0f What's ROCm, and how does it compare with CUDA?\n\n- ROCm (Radeon Open Compute) is AMD\u2019s open-source software stack for enabling GPU (and compute) acceleration on GPUs, similar to CUDA for NVIDIA. It includes driver support, compilers, libraries, and runtimes for frameworks like PyTorch, TensorFlow, ONNX and more. ([rocmdocs.amd.com](https://rocmdocs.amd.com/projects/radeon/en/latest/index.html?utm_source=openai))  \n- Performance for inference (running models) has come forward strongly in recent versions. For example, ROCm 6.2+ adds enhanced FP8 support and boosts inference/training on AMD\u2019s \u201cInstinct\u201d and consumer RDNA/GPU lines. ([sdxcentral.com](https://www.sdxcentral.com/news/amd-launches-rocm-62-adds-fp8-support-and-enhanced-ai-training-and-inferencing-capabilities/?utm_source=openai))  \n- On the flip side, in benchmarks and comparisons, CUDA + Nvidia GPUs still hold advantages in certain areas like large-scale training, support for clipping-edge features, optimized libraries (cuDNN, TensorRT, etc.), and multi-GPU scaling via NVLink. ROCm is catching up but has gaps. ([compute-market.com](https://www.compute-market.com/blog/amd-vs-nvidia-for-ai-2026?utm_source=openai))  \n\n---\n\n## \ud83d\udda5\ufe0f How well do models run with AMD GPUs?\n\nDepends heavily on **which GPU/APU, how much VRAM**, and the **specific model size** (parameter count), precision, etc.\n\nHere are some specifics:\n\n- For **inference** (running existing models), many users run Llama, Mistral, Qwen3 etc., on RDNA3 and RDNA4 GPUs (24-16GB+ VRAM) with ROCm 7.x, with performance \u201crivalling an RTX 3090\u201d in some cases. ([llmhardware.io](https://llmhardware.io/guides/amd-rocm-llm-guide?utm_source=openai))  \n- It\u2019s more about **suitability** for model size. Smaller models (e.g. 7-13B parameters) map well on \u201cmid-to-high end\u201d RDNA3/4 or newer APUs. But large models\u201470B, 100B+\u2014either need much more VRAM or creative techniques like quantization, offloading, or sharding. ([localaiops.com](https://localaiops.com/posts/running-local-llms-on-amd-gpus-with-rocm-and-ollama/?utm_source=openai))  \n\n---\n\n## \u26a0\ufe0f Where AMD + ROCm still have limitations\n\n- **Training large models**: For sizable training tasks (many parameters, full precision or mixed precision), NVIDIA + CUDA often still have better tooling, more mature optimizations, and more stable multi-GPU support. ROCm is improving but lags in some of the newer features. ([compute-market.com](https://www.compute-market.com/blog/amd-vs-nvidia-for-ai-2026?utm_source=openai))  \n- **Support for certain advanced features**: Some quantization formats, attention/kernel optimizations (FlashAttention, etc.), or exotic hardware features can have partial or delayed ROCm support. ([llmhardware.io](https://llmhardware.io/guides/amd-rocm-llm-guide?utm_source=openai))  \n- **Software ecosystem & debugging**: Many AI tools still assume CUDA first. Installing/setting up ROCm, finding prebuilt binaries for all tools, or finding detailed bug reports may be harder. ([kunalganglani.com](https://www.kunalganglani.com/blog/rocm-consumer-gpu-cuda-alternative-2026?utm_source=openai))  \n\n---\n\n## \ud83d\udd0d What this means for a system like the Framework PC with an AMD GPU\n\nIf your Framework PC is using something like an AMD Ryzen AI Max (Strix Halo) APU or a Radeon GPU (likely RDNA3 or newer for \u201cAI-friendly\u201d versions):\n\n- You\u2019ll *definitely* be able to run a good number of popular models locally, especially smaller ones, inference tasks, image generation, etc.\n- You\u2019ll need to ensure ROCm support for your specific GPU/APU, the right drivers, up-to-date software (OS, kernel), and possibly use quantization/co-processor-friendly tricks to squeeze more out.\n- For cutting-edge training of massive models, or when ultra-low latency + highest throughput is needed, NVIDIA + CUDA setups may still offer better ROI.\n\n---\n\nIf you like, I can pull up how a **specific model** (e.g. Llama 3, or Stable Diffusion, or a 70B LLM) performs on an AMD GPU/APU similar to what Framework is offering, so you can see concrete trade-offs."
    },
    {
        "role": "user",
        "content": "give%20me%20some%20good%20models%20i%20can%20run%20on%20this%20hardware%20for%20coding%20(especially%20python)"
    },
    {
        "role": "assistant",
        "content": "If you're running on a system with an AMD GPU or APU (especially something high-end like an RDNA3/4-based card or a Ryzen AI Max device), there are several excellent open-source coding models that balance performance and capability. Here are some of the best options, along with hardware advice and how they stack up for **Python coding** tasks like autocompletion, refactoring, debugging, and generating multi-file modules.\n\n---\n\n## \ud83e\udde0 Top Models for Local Coding (Python-Focused)\n\n| Model | Size / Parameters | Strengths for Coding | Hardware Needs / Fit for AMD GPU |\n|---|---|---|---|\n| **Qwen3-Coder-Next** | ~80B MoE (activates \u2248 3B per pass) | Excellent for coding agents: very strong at debugging, multi-step reasoning, SWE-Bench Verified performance. ([byteiota.com](https://byteiota.com/qwen3-coder-next-local-deployment/?utm_source=openai)) | Because only a small fraction of the model is active per inference, it works decently well even with moderate VRAM. You\u2019ll want 12\u201320 GB VRAM to get a smooth experience. |\n| **Qwen 3.5 Coder** | 7B / 14B / 32B | Good all-around for Python, supports completions, multiple file contexts. ([developersdigest.tech](https://www.developersdigest.tech/guides/run-ai-models-locally?utm_source=openai)) | The 7B fits on ~8GB GPU well, 14B starts getting more demanding. Dense variants (like 32B) need ~20GB+; shared-VRAM AMD Strix Halo setups may manage some of these. |\n| **DeepSeek Coder V2 / Lite** | MoE-based, \u201cLite\u201d variants are ~16B or less | Great mix of speed and coding quality; strong on Python, Go, Rust. ([promptquorum.com](https://www.promptquorum.com/local-llms/best-local-llms-for-coding?utm_source=openai)) | The Lite/MoE variants are good fits for 16GB of VRAM; denser variants need more. For daily work, something under 20-24 GB tends to give good latency. |\n| **Llama 3.3 / CodeLlama** | 70B / 34B / smaller | Best large local models; great long-context performance and code structure understanding. CodeLlama-Python optimized for Python tasks. ([localaimaster.com](https://localaimaster.com/models/best-local-ai-coding-models?utm_source=openai)) | 70B needs a lot (24-32+ GB VRAM) to run well; smaller variants 13B-34B are more realistic on modern AMD cards with shared VRAM / high memory headroom. |\n| **GLM-4.7-Flash** | 30B with MoE (~3B active per pass) | Strong for agentic workflows, troubleshooting, and tool-calling. High benchmarks on coding benchmarks. ([pondhouse-data.com](https://www.pondhouse-data.com/blog/glm-4-7-flash-local-opencode?utm_source=openai)) | Quantized versions (e.g. Q4) can run on ~16GB GPU, a good match for high-end consumer AMD GPUs or shared GPU/APU setups with sufficient allocation. |\n| **CodeLlama-7B-Python (GGUF quantized)** | ~7B params | Specially fine-tuned for Python: great for code completion, analysis, bug fixes. Multiple quantization options available so you can trade precision for speed. ([promptlayer.com](https://www.promptlayer.com/models/codellama-7b-python-gguf/?utm_source=openai)) | Excellent fit for 6-12GB VRAM setups. With quantization (especially 4-bit or similar), performance gets very usable. |\n\n---\n\n## \ud83d\udd0d What to Aim for with Your Hardware\n\nIf your machine is something like the Framework Desktop with an AMD GPU/APU, here\u2019s a guide for what size of model you\u2019ll likely be able to use comfortably:\n\n- **Modern AMD with 20\u201340GB VRAM or in shared-VRAM high spec**: You can handle 27B dense or MoE models like Qwen-3.6-27B, GLM-4.7-Flash, etc., with reasonable latency (~10-20 tokens/sec depending on inference engine and quantization).  \n- **Mid-range AMD (12-16GB VRAM or shared high allocation)**: Models like Qwen-3.5-Coder (14B), DeepSeek Coder Lite (~16B MoE), or GLM-4.7-Flash with quantization are sensible.  \n- **Lower VRAM budget (<10-12GB)**: Go for 7B-13B class models, quantization (4-bit or similar), smaller context lengths, and consider CPU fallback for parts if needed. Models like CodeLlama-Python 7B, Phi-4 Mini (~3.8B), Qwen 8B are good here.  \n\n---\n\n## \u26a0\ufe0f Things to Watch Out For\n\n- **Quantization Compatibility**: Not all quantization formats are equally optimized on AMD hardware. For example, \u201cK-quant\u201d variants or certain GGUF formats might run slower unless ROCm + HIP / Vulkan backends are well-configured. ([craftrigs.com](https://craftrigs.com/tools/gpu-compatibility-matrix-local-llms/?utm_source=openai))  \n- **Shared VRAM**: On Ryzen AI APUs or setups where graphics memory is \u201cshared\u201d with system memory, allocating more to the GPU helps\u2014but leads to trade-offs in what else the system can do.  \n- **Inference Engine / Runtime**: Use tools like **Ollama**, **llama.cpp (HIP backend)**, **ZINC (Vulkan shaders)**, etc. These often provide better performance on AMD than vanilla CUDA-centric setups. ([github.com](https://github.com/zolotukhin/zinc?utm_source=openai))  \n- **Context Window / Mult-File Editing**: Big context windows (8-16K tokens and beyond) are great but scale VRAM use significantly. Keep tabs on context-length costs.  \n\n---\n\nIf you tell me the exact specs of your Framework PC (GPU/APU model, how much VRAM/shared memory you can allocate, CPU cores, etc.), I can suggest 2-3 models that\u2019ll give you the **best coding experience** for Python."
    },
    {
        "role": "user",
        "content": "is%20all%20ram%20on%20the%20framework%20pc%20available%20to%20the%20gpu%3F"
    },
    {
        "role": "assistant",
        "content": "Here\u2019s what I found out about the Framework Desktop and how its RAM works with the GPU / iGPU. TL;DR: yes, pretty much all the system RAM *can* be used by the GPU\u2014but with caveats.\n\n---\n\n## \ud83d\udda5\ufe0f How the Framework Desktop\u2019s Memory Architecture Works\n\n- The Framework Desktop uses a unified memory design (UMA/UPM) because it\u2019s built around an APU (AMD\u2019s Ryzen AI Max / Strix Halo). There is **no separate, dedicated VRAM** like a discrete graphics card has. All graphics memory comes from the same pool of system RAM. ([framework.wiki](https://framework.wiki/docs/computers/desktop/amd-ryzen-ai-max/?utm_source=openai))  \n- The RAM is soldered LPDDR5x (\u201cnon-upgradeable\u201d) and shared across CPU, iGPU, and other subsystems. ([framework.wiki](https://framework.wiki/docs/computers/desktop/amd-ryzen-ai-max/?utm_source=openai))  \n\n---\n\n## \ud83d\udcca How Much RAM the GPU Can Use\n\n- According to user reports and BIOS/driver behaviors, you can dynamically allocate how much RAM the GPU needs. In practice, under Windows, people say you can allocate \u201cup to system RAM minus ~8 GB\u201d to the iGPU. On Linux, possibly even more, when running headless or without a heavy desktop environment. ([community.frame.work](https://community.frame.work/t/igpu-vram-how-much-can-be-assigned/73081?utm_source=openai))  \n- For example, with a 128 GB total memory system: users noted that roughly 120\u2013122 GB of it could be set aside for GPU use (with the rest for other OS overhead, etc.). ([community.frame.work](https://community.frame.work/t/igpu-vram-how-much-can-be-assigned/73081?utm_source=openai))  \n\n---\n\n## \u26a0\ufe0f Caveats & Practical Limitations\n\n- Even though *technically* most of the RAM *can* be used by the GPU, doing so means less RAM is left over for other tasks\u2014especially CPU-heavy ones. If you allocate too much to the GPU, things like multitasking, compilation, running other apps might get slow. ([community.frame.work](https://community.frame.work/t/igpu-vram-how-much-can-be-assigned/73081?utm_source=openai))  \n- Performance can drop if you rely on RAM rather than faster, cache-optimized memory for graphics workloads. RAM bandwidth, latency, and caching behavior are not as ideal as GPU-dedicated memory. ([community.frame.work](https://community.frame.work/t/igpu-vram-how-much-can-be-assigned/73081?utm_source=openai))  \n- Some of the memory allocated to the GPU is reserved *at boot time* (not dynamically changed often at runtime, though BIOS / driver tools might give some flexibility). That means you need to decide ahead of use how much you want available for GPU-intensive tasks. ([community.frame.work](https://community.frame.work/t/igpu-vram-how-much-can-be-assigned/73081?utm_source=openai))  \n\n---\n\nIf you want, I can pull up specific memory-allocation settings (BIOS / driver) for your Framework device so you know how to adjust VRAM allotment for Linux or Windows."
    },
    {
        "role": "user",
        "content": "check%20prices%20on%20ebay%20for%20used%2Frefurb%20framework%20desktops"
    },
    {
        "role": "assistant",
        "content": "Here are some recent used/refurbished listings for the Framework Desktop on eBay, including what sold and what\u2019s currently available, so you can get an idea of typical pricing:\n\n---\n\n## \ud83d\udcb8 Recent Sold Listings\n\n| Configuration | Condition / Notes | Sold Price | When |\n|---|---|---|---|\n| Framework Desktop Max+ 395, **64 GB RAM**, factory setup | \u201cOpen box\u201d, used only a few days | **$2,000** | April 27, 2026 ([ebay.com](https://www.ebay.com/itm/257446817669?itmmeta=01KNMHQDP97ZG12WPES7MV31XE&itmprp=enc%3AAQALAAAA4DKQclQvzFwZQpmMrsO4LurYTyaxBH9GJhFTPHSgeA5c6%2FTdlJ%2FU%2BFAWMyLqQm72bpPupPjG9XFq02gnTjLXK2rk76MKLbCOBuPKB4FfyI6%2B%2BKvd6dnh25sd%2BKLHkuRav7yI5aKHWnsugMBzzUi9ad8Vu2PRVr6A4iKR%2FO3cIs4B%2Fn%2FFPc1CLBZ1sVpOan30i9yhNG0JmGZrbUN0pNQFIf997ktZopioFkNsfF6v2pkYkBpXxENFCi73ArHA4iC8bzFUHlIaYv%2BYuF%2FsrDX1NABMBCz0p0ioWg8tEsJ3N2QV%7Ctkp%3ABk9SR6rb3ZGtZw&utm_source=openai)) |\n\n---\n\n## \ud83d\udd0d Current Listings\n\n| Configuration | Status / Notes | Asking Price |\n|---|---|---|\n| Framework Desktop Max+ 395, **128 GB RAM** (brand new) | Fully new units with full specs | **$2,699.99** ([ebay.com](https://www.ebay.com/itm/257100082494?utm_source=openai)) |\n| Other \u201cused/pre-owned\u201d Framework Desktop / tower-types | May include lower RAM or partial specs | Various\u2014few high-spec units listed near $2,000\u2013$2,500 ([ebay.com](https://www.ebay.com/shop/framework-computer?_nkw=framework+computer&utm_source=openai)) |\n| Framework Desktop Max+ 395, 16-core / 128 GB unit (reseller stock) | Used for testing/benchmarks by hobbyists/benchers | **$2,799.99** ([ubbcentral.com](https://www.ubbcentral.com/store/item/Framework-Desktop-Ryzen-AI-Max-395-16-Core-128GB-LPDDR5x-1TB-NVMe_326791788137.html?utm_source=openai)) |\n\n---\n\n## \ud83d\udcc9 Market Trend and What to Expect\n\n- For a high-end spec (128 GB RAM, top-tier CPU/APU), **brand new** units sell for approx **$2,700\u2013$2,800**.  \n- Refurbished/open-box or lightly used models with around **64 GB** of RAM show up at approx **$2,000**.  \n- If you find a used one with 32\u201364 GB, or maybe parts replaced, expect cheaper, but strong units (especially with 128 GB) almost always stay high-priced.\n\n---\n\nIf you tell me your target budget or the spec (RAM, CPU/APU, storage) you're looking for, I can scan eBay now and pull up a few options that fit well."
    },
    {
        "role": "user",
        "content": "what%20are%20the%20framework%20desktop%20pc%20models%3F"
    },
    {
        "role": "assistant",
        "content": "Here are the current models and specifications of the Framework Desktop PC line (as of mid-2025). The Desktop comes in a few distinct configurations\u2014mainly distinguished by processor, graphics, and RAM capacity. All share the same core design and many features.\n\n---\n\n## \ud83d\udd27 Framework Desktop Models & Configurations\n\n| Model Name | CPU / APU (soldered) | GPU (Integrated) | RAM (soldered) | Starting Price* |\n|---|---|---|---|---|\n| Max 385 | AMD Ryzen AI Max 385 \u2014 8 cores / 16 threads ([techspot.com](https://www.techspot.com/news/106934-framework-first-ever-desktop-pc-packs-amd-ryzen.html?utm_source=openai)) | Radeon 8050S (32 compute units) ([tweakers.net](https://tweakers.net/nieuws/232270/frameworks-eerste-desktop-pc-heeft-een-ryzen-ai-max-cpu-en-maximaal-128gb-ram.html?utm_source=openai)) | 32 GB LPDDR5x-8000 ([tomsguide.com](https://www.tomsguide.com/computing/mini-pcs/framework-desktop-review?utm_source=openai)) | About **US$1,099** ([pcworld.com](https://www.pcworld.com/article/2866400/framework-desktop-review.html?utm_source=openai)) |\n| Max+ 395 | AMD Ryzen AI Max+ 395 \u2014 16 cores / 32 threads ([techspot.com](https://www.techspot.com/news/106934-framework-first-ever-desktop-pc-packs-amd-ryzen.html?utm_source=openai)) | Radeon 8060S (40 compute units) ([tweakers.net](https://tweakers.net/nieuws/232270/frameworks-eerste-desktop-pc-heeft-een-ryzen-ai-max-cpu-en-maximaal-128gb-ram.html?utm_source=openai)) | 64 GB or 128 GB LPDDR5x-8000 ([techspot.com](https://www.techspot.com/news/106934-framework-first-ever-desktop-pc-packs-amd-ryzen.html?utm_source=openai)) | ~$1,599 for 64 GB; ~$1,999 for 128 GB ([tomsguide.com](https://www.tomsguide.com/computing/mini-pcs/framework-desktop-review?utm_source=openai)) |\n\n\\* *Prices before add-ons like more storage, extra cooling, or decorative tiles.*\n\n---\n\n## \ud83c\udf10 Shared Features Across All Models\n\n- **Form Factor & Design**: Mini-ITX size, 4.5-liter chassis. Uniform layout, heatsink, and power supply across all models. ([framewiki.net](https://framewiki.net/products/framework-desktop?utm_source=openai))  \n- **Power Profile**: All models run at ~120W sustained, with boost up to ~140W under higher loads. ([phoronix.com](https://www.phoronix.com/review/framework-desktop-linux?utm_source=openai))  \n- **Memory**: LPDDR5x-8000, soldered (non-upgradable) for bandwidth reasons. ([tomshardware.com](https://www.tomshardware.com/desktops/gaming-pcs/framework-desktop-review?utm_source=openai))  \n- **Storage Options**: Two M.2 NVMe SSD slots, from ~500 GB up to 8 TB (depending on what Framework offers). ([pcworld.com](https://www.pcworld.com/article/2866400/framework-desktop-review.html?utm_source=openai))  \n- **Connectivity & Ports**: Rear IO includes USB-A, USB-C / USB4, DisplayPort 2.1, HDMI 2.1, and a 5 Gigabit Ethernet port. Front has slots for expansion cards (USB-C, USB-A, etc.). ([pcworld.com](https://www.pcworld.com/article/2866400/framework-desktop-review.html?utm_source=openai))  \n- **Graphics & AI Performance**: Integrated GPU is capable of gaming (1080p to 1440p depending on setting) and AI workloads; includes an NPU (~50 TOPS) for inference tasks. ([tomsguide.com](https://www.tomsguide.com/computing/mini-pcs/framework-desktop-review?utm_source=openai))  \n\n---\n\nIf you want, I can compare these to the Framework Laptop line or see if there have been updated models past 2025."
    }
]