Fundamental market or economic problem — can't be fixed by changing execution. Don't invest further.
{ "market_score": 5, "tam_usd_b": 0.34, "sam_usd_m": 21.1, "som_usd_m": 0.38, "tam_methodology": "Bottom-up: (1) Global developers ~28.7M; assume 20% will actively run local LLMs in next 12–24 months => 5.7M. Of those, 25% would pay for better runtime/UX/perf => 1.43M seats at ~$144/yr => ~$206M. (2) AI hobbyists willing to pay: ~1.0–1.5M at ~$60/yr => $60–$90M (midpoint $75M). (3) SMB/team licenses: ~150k seats at ~$240/yr => ~$36M. Sum midpoint ≈ $317–$347M; we use $340M TAM to reflect upside on Windows/Linux adoption.", "competitors": [ { "name": "Ollama", "price": "$0 (open-source CLI/server); enterprise features rumored", "revenue_est": "$0–$5M ARR (est., mostly indirect/enterprise pilots)", "strength": "Extremely simple developer UX and massive community/model registry momentum on macOS and Linux.", "weakness": "General-purpose runtime; no deep specialization for MoE expert streaming from SSD on low-RAM Macs." }, { "name": "llama.cpp", "price": "$0 (open-source library/CLI)", "revenue_est": "$0 (OSS core; commercial forks exist)", "strength": "Highly optimized Metal kernels and ubiquitous GGUF ecosystem with very wide hardware coverage.", "weakness": "Primitive MoE scheduling and I/O prefetching; UX and packaging for teams is minimal." }, { "name": "LM Studio", "price": "Free; Pro/Teams ~$15–$20/user/mo (est.)", "revenue_est": "$1–$3M ARR (est., based on download base and plausible conversion)", "strength": "Best-in-class desktop UX for local models with curated downloads and auto-setup on Apple Silicon.", "weakness": "Focus on GUI breadth over novel runtimes; MoE/SSD streaming for low-memory edge cases not a core differentiator yet." }, { "name": "MLC LLM", "price": "$0 (open-source compiler/runtime)", "revenue_est": "$0 (research/OSS project)", "strength": "Ahead on Apple Silicon codegen/compilation (TVM/Metal); deep performance chops.", "weakness": "Complex setup; not a turnkey product for prosumers, and limited productization around MoE disk streaming." }, { "name": "Apple MLX / Core ML + Metal", "price": "Bundled $0 (platform SDKs)", "revenue_est": "$0 (platform line item)", "strength": "First-party access to Metal/AMX and distribution via Xcode; could bundle local LLM
# Competitive Intelligence Report: Local LLM Platforms Competing with SSD‑Streamed MoE on Apple Silicon The niche this report examines is the emerging market for high‑end local large‑language‑model (LLM) inference on consumer hardware, specifically the ability to run models such as **Qwen3.6‑35B‑A3B** or DeepSeek‑4 class Mixture‑of‑Experts (MoE) models on devices like a 16 GB M1 Pro MacBook using SSD‑streamed weights rather than traditional full RAM loading.[2][3][10] The ds4 project for DeepSeek‑4 Flash and Pro demonstrates that, with asymmetric quantization and SSD expert streaming via `mmap`, it is already possible to run 30–32 GB MoE models on systems with materially less RAM by turning RAM into a cache for hot experts and treating SSD as an extension of the memory hierarchy.[3][10] Against this backdrop, several local LLM platforms—Ollama, LM Studio, Jan.ai, GPT4All (by Nomic AI), and Private LLM for Apple devices—constitute the most dangerous competitors, given their traction, funding, and positioning as default choices for developers and privacy‑sensitive users who want local inference. This report provides deep competitive intelligence on these players, benchmarks pricing and willingness to pay in this niche, and identifies market gaps that an SSD‑streamed MoE product for Qwen3.6‑35B‑A3B on Apple Silicon could exploit. ## 1. Market Context: Local LLMs, MoE Streaming, and Apple Silicon ### 1.1 The Business Idea in Technical and Market Terms The business idea underlying this research—“Show HN: Qwen3.6‑35B‑A3B on a 16 GB M1 Pro with SSD‑streamed MoE”—sits at the intersection of three trends: the rise of open‑weight frontier‑adjacent models such as Qwen3.6, the maturation of MoE architectures, and the rapid growth of local LLM tooling for laptops and workstations.[2][3][10] Qwen3.6‑35B‑A3B is part of the Qwen 3.6 line, which emphasizes agentic coding and long‑context reasoning, and the maintainers explicitly recommend very large output lengths, up to 32,768 tokens for most queries and as high as 81,920 tokens for complex math and programming tasks.[2] That level of context, combined with 35B parameters and an A3B architecture, pushes beyond what straightforward in‑RAM loading can handle on a 16 GB M1 Pro, especially if one wants interactive performance rather than batch‑style offline generation.[2][3] MoE architectures, particularly routed experts with quantization, offer a path to reconcile model size and hardware limits by activating only a subset of experts per token and streaming inactive experts from storage as needed.[3][10] The ds4 project by antirez (Salvatore Sanfilippo) shows this concretely for DeepSeek‑4 models: SSD streaming allows the available RAM to be treated not as a hard cutoff, but as a continuum of speed levels, with routed MoE experts kept in an in‑memory cache while cold experts are streamed from disk using `mmap`.[3][10] On Windows, ds4 replicates a technique first tested on Apple systems by Daniel Isaac, mapping the model file directly into virtual memory so the operating system page cache becomes an “expert cache manager,” caching hot experts in RAM and streaming cold ones from SSD at roughly 0.5–1.5 GB/s depending on hardware.[3] In experiments, this allowed a 32 GB Q4‑quantized MoE model to run on a 28 GB system at 2.5–4 tokens per second, whereas a non‑streamed configuration would have been impossible.[3] For Apple Silicon, tooling such as `llama.cpp` and `llama-cpp-python` already supports Metal GPU offload and GGUF quantized models, enabling small to mid‑size models like CodeLlama‑7B to run efficiently on macOS by specifying `GGML_METAL=on` and setting `--n_gpu_layers` for GPU acceleration.[8] However, mainstream local LLM platforms like Ollama and LM Studio currently focus more on static GGUF quantized models and unified‑memory loading rather than deep MoE‑specific SSD streaming on constrained RAM.[4][34] This creates an
# Market Sizing and Risk Analysis for “Show HN: Qwen3.6‑35B‑A3B on a 16 GB M1 Pro with SSD‑Streamed MoE” This report analyzes the commercial potential and risk profile of a business built around enabling Qwen3.6‑35B‑A3B, a large mixture‑of‑experts (MoE) model, to run locally on a 16 GB M1 Pro MacBook using SSD‑based streaming and Apple Metal‑optimized inference, positioned as a “Show HN” open‑source project that may evolve into a product or company.[1][7][8][9] The core idea is to exploit SSD streaming of MoE expert weights so that a model which nominally requires more than 25 GB of VRAM can be executed on consumer‑grade Apple Silicon laptops with only 16 GB of unified memory, thereby unlocking high‑end reasoning and coding capabilities for developers and power users who prefer or require local inference.[1][8][9][10] Using available market data on generative AI software, professional development spending, developer populations, and Apple Silicon hardware, this report constructs scenario‑based bottom‑up estimates of total addressable market (TAM), serviceable addressable market (SAM), and serviceable obtainable market (SOM), while explicitly flagging where empirical data is missing and assumptions are required.[14][15][16][17][18] It also assesses historical evidence of failure or pivot among comparable local‑LLM tooling projects, examines the regulatory and legal landscape for on‑device AI and model distribution, and surveys recent funding and competitive signals in adjacent domains such as open‑model agents, local inference platforms, and on‑device foundation models.[11][12][14][19] Throughout, the analysis emphasizes that this niche—SSD‑streamed MoE inference for large open models on consumer Apple Silicon—is very young, with limited direct precedent, and hence both the upside and the risk are unusually high, especially regarding technical execution, developer adoption, and competitive moves by larger incumbents.[7][8][9][19] ## 1. Business Concept and Context ### 1.1 Defining the Proposed Business Idea The business idea can be summarized as building and potentially commercializing an open‑source, highly optimized inference stack that runs Qwen3.6‑35B‑A3B locally on a MacBook Pro M1 Pro with 16 GB of unified memory by streaming MoE expert weights from a fast NVMe SSD, leveraging Apple’s Metal GPU APIs and modern quantization and caching techniques.[1][7][8][9][19] Qwen3.6‑35B‑A3B is a large, advanced generative model developed by Alibaba’s Qwen team, characterized by strong performance on coding and reasoning tasks, a default context length of 262,144 tokens (with mechanisms to push beyond 1 million tokens via RoPE scaling), and a “thinking mode” that emits internal chain‑of‑thought content before final answers.[9] The Qwen3.6‑35B‑A3B variant is a mixture‑of‑experts model, meaning it contains multiple expert subnetworks whose weights contribute significantly to the model’s total memory footprint, which is why typical GPU deployments require more than 25 GB of VRAM and are not directly compatible with a 16 GB MacBook Pro M1 Pro.[1][8][9] WillItRun.ai documents that Qwen 3.5 35B in A3B quantization requires approximately 25.3 GB of VRAM and therefore cannot run natively on a MacBook Pro M1 Pro 16 GB, which offers about 11.5 GB of accessible GPU memory, highlighting the hardware constraint that this project aims to circumvent via SSD streaming and software ingenuity.[1] The SSD‑streamed MoE concept builds on ideas articulated by MindStudio in its description of “SSD Streaming for AI Models,” where model weights—especially those of experts in a MoE architecture—are stored on an NVMe SSD rather than fully in RAM, and only the active experts are loaded into memory on demand at inference time.[8] In MindStudio’s “Dwarf Star” design, non‑expert components such as attention layers, layer normalization, and routing weights remain in RAM, while expert weights, which account for much of the model’s size, are offloaded to disk and fetched as needed, enabling significantly reduced RAM requirements at the cost of some inference speed.[8] This technique allows developers to run larger models than their system RAM would otherwise allow, provided they have a sufficiently fast SSD, ideally PCIe 4.0 or better with sequential read speeds around 5,000–7,000 MB/s to support effective prefetching.[8] Applied to Qwen3.6‑35B‑A3B on a 16 GB M1 Pro, the idea is that expert weights for the MoE model can be streamed from SSD, while the routing, attention, and other shared components reside within the constrained unified memory, thus making local inference feasible despite the model’s large parameter count and context length.[1][8][9] The project URL referenced in the query points to a GitHub repository related to ds4, an inference engine originally created for DeepSeek V4 Flash and optimized for Apple’s Metal API, with ds4‑webui presented as a minimal frontend and Pinokio launcher providing a browser‑based interface to a Metal‑only inference server.[6][7] The ds4‑webui repository describes itself as “a Pinokio launcher and standalone browser UI for antirez/ds4, a narrow Metal‑only inference engine for DeepSeek V4 Flash,” emphasizing that it is tailored to on‑device usage and Apple Silicon, mirroring the focus of this Qwen3.6 business idea.[7] This ecosystem indicates that there is already a technical and community foundation for Metal‑optimized inference engines and HN‑style show‑and‑tell projects that subsequently evolve into more polished tools or platforms.[6][7][11] The proposed business would likely follow a similar pattern: launching as an open‑source “Show HN” demonstration proving that Qwen3.6‑35B‑A3B can run on a 16 GB M1 Pro via SSD‑streamed MoE, then exploring monetization such as pro distributions, managed installers, enhanced agent capabilities, or enterprise support for teams standardizing on local Qwen tooling.[6][7][9][11][12] ### 1.2 Key Differentiators Versus Existing Local LLM Tools To understand market size and risk, it is crucial to distinguish this idea from existing local LLM platforms that already support Apple Silicon but generally target smaller models or different usage patterns.[10][11][12] WillItRun.ai, for example, maintains rankings of the “Best Local LLMs for MacBook Pro M1 Pro 16GB,” recommending lighter Qwen models such as Qwen 3.5 9B, Qwen 3 8B, and Qwen 3.5 4B as best picks for coding, chat, and writing on this hardware, without attempting to run 35B‑scale models due to memory constraints.[10] This shows that today’s mainstream experience for a developer on a 16 GB M1 Pro is to use 4–9B parameter models via platforms like Ollama, LM Studio, or Jan, rather than pushing the hardware to run 35B MoE models with multi‑hundred‑thousand‑token context windows.[4][10][11][12] LM Studio positions itself as “Bionic, an agent for open models, natively local, built for creativity, work, and code,” emphasizing a general‑purpose agent layer that can orchestrate various open models but not necessarily focusing on extreme model size or SSD streaming innovations on constrained laptops.[11] Jan.ai markets itself as an “open‑source alternative to ChatGPT” that can run open‑source AI models locally or connect to cloud models like GPT and Claude, again focusing on usability and hybrid local‑cloud connectivity more than on enabling very large MoE models on resource‑limited hardware.[12] By contrast, the SSD‑streamed Qwen3.6‑35B‑A3B idea targets a more technically advanced niche: developers, researchers, and power users who want near‑frontier performance and very long context lengths but either prefer local inference for privacy or cost reasons, or simply enjoy pushing Apple Silicon hardware to its limits.[1][8][9][10] In effect, this is closer in spirit to ds4 and Dwarf Star than to LM Studio or Jan.ai, prioritizing highly optimized inference pipelines, Metal acceleration, and streaming architectures, with user‑facing UX potentially arriving later as the technology stabilizes.[7][8][11][12] Such a product could be differentiated by supporting “thinking mode” in Qwen3.6, which preserves reasoning context in historical messages and leverages extended context windows to improve agentic coding workflows and repository‑level reasoning, something explicitly highlighted by Qwen3.6 documentation as a key capability.[9] For users engaged in intensive coding, front‑end workflows, and simulation tasks in a browser‑based environment, a local Qwen3.6‑35B‑A3B deployment on a MacBook could be highly attractive, especially if coupled with agent frameworks and caching strategies tuned for Apple Silicon.[2][9][11][19] ### 1.3 Relationship to Apple’s On‑Device AI Strategy The idea also exists within a broader strategic environment in which Apple itself is investing heavily in on‑device foundation models optimized for Apple Silicon, a trend that both validates the importance of on‑device LLMs and creates potential competitive risk.[19] Apple’s machine learning research updates describe an on‑device foundation model optimized for efficiency and tailored specifically for Apple Silicon, designed to enable low‑latency inference with minimal resource usage, along with larger server‑side models that can complement on‑device capabilities.[19] This suggests that Apple believes there is substantial value in keeping at least part of AI inference local, leveraging Apple Silicon’s GPU and Neural Engine to offer private, low‑latency experiences that do not rely entirely on the cloud.[13][19] At the same time, Apple’s enthusiasm for on‑device models raises the possibility that future macOS or Xcode versions could ship built‑in LLMs and agent layers, reducing the need for third‑party local inference stacks in some scenarios, or imposing new platform constraints that third‑party tooling must navigate.[13][19] For the Qwen3.6‑35B‑A3B SSD‑streamed MoE project, Apple’s direction is a mixed signal: it indicates clear user demand and platform support for on‑device AI, yet it suggests that long‑term differentiation may depend on supporting open models, highly specialized workflows (such as large‑context coding agents), and cross‑platform capabilities that Apple’s own models may not prioritize.[9][11][19] Apple’s documentation of Macs with Apple Silicon highlights a wide range of models, from the MacBook Air to the Mac Pro, that share the same fundamental architecture and benefit from unified memory and dedicated accelerators, meaning that any Metal‑optimized SSD‑streamed inference stack built for the M1 Pro could potentially be extended to other Apple Silicon devices, broadening the hardware base.[13][19] However, the business would need to navigate Apple’s evolving policies on Intel‑based app support via Rosetta, which is slated to end in a future macOS version, and ensure that the tooling is compiled natively for Apple Silicon rather than relying on Intel translation layers that might compromise performance or future compatibility.[5][13] Overall, the business idea aligns with the platform trajectory toward on‑device AI but must anticipate a world in which Apple’s own foundation models and APIs become standard components and third‑party tools compete by offering more model choice, control, and advanced technical features such as SSD‑streamed MoE.[9][13][19] ## 2. Technology and Use‑Case Landscape ### 2.1 Qwen3.6‑35B‑A3B Capabilities and Requirements Qwen3.6‑35B‑A3B occupies the high end of the open‑model landscape, aimed at demanding tasks such as agentic coding, long‑horizon reasoning, and multimodal understanding.[2][9] The Hugging Face model card for Qwen3.6‑35B‑A3B describes this release as delivering substantial upgrades in agentic coding, enabling more fluent handling of frontend workflows and repository‑level reasoning, which is particularly relevant for developers working with complex codebases on their local machines.[9] It also emphasizes “thinking preservation,” whereby the model operates in thinking mode by default, producing `<think>…</think>` content to represent internal reasoning before emitting final responses; users can choose to preserve or suppress this thinking content depending on task requirements, and the model is designed to make good use of extended context for iterative development and agent scenarios.[9] The model’s default context length is 262,144 tokens, with support for long‑context techniques such as RoPE scaling to handle tasks where total input and output length can exceed this limit, potentially approaching 1,010,000 tokens in specialized configurations.[9] Such a context is valuable for developers who wish to load entire repositories, documentation sets, or long‑running conversational histories into the model, but it also compounds memory and compute requirements, making efficient KV cache management and streaming strategies important.[8][9] From a hardware perspective, Qwen3.6‑35B‑A3B is demanding even when quantized, and typical deployments rely on multi‑GPU configurations with large VRAM or cloud instances with ample memory.[1][9] WillItRun.ai notes that the prior Qwen 3.5 35B A3B quantization requires 25.3 GB of VRAM, exceeding the practical GPU memory available on MacBook Pro M1 Pro 16 GB, which only exposes around 11.5 GB VRAM to such workloads.[1] While Qwen3.6 has been optimized relative to earlier Qwen models, including improvements in KV cache utilization and efficiency in thinking and non‑thinking modes, the model card still advises maintaining at least a 128K token context window to preserve thinking capabilities in complex tasks, implying a substantial memory footprint across weights and cache.[9] For video understanding, Qwen3.6 suggests adjusting video preprocessor parameters to support up to 224k video tokens for hour‑scale videos, again underscoring the model’s design target of long‑horizon multimodal tasks that are far beyond the scope of lightweight local assistants.[9] These capabilities translate directly into the business proposition: by making Qwen3.6‑35B‑A3B accessible on a 16 GB M1 Pro via SSD‑streamed MoE, the project would bring near‑frontier coding and reasoning power directly onto mainstream developer laptops, enabling workflows such as repository‑scale code review, full‑history agentic debugging, and long‑context simulation or design tasks without reliance on cloud APIs.[1][8][9][10] This could be particularly compelling for developers working with sensitive codebases who are constrained by confidentiality policies
# Organic Demand Signals For “Show HN: Qwen3.6‑35B‑A3B on a 16 GB M1 Pro with SSD‑Streamed MoE” The business idea under examination is a developer‑focused solution that makes it practical to run Qwen‑style hybrid Mixture‑of‑Experts (MoE) models, specifically Qwen3.6‑35B‑A3B, on constrained Apple Silicon hardware such as a 16 GB M1 Pro MacBook Pro by aggressively streaming inactive expert weights from SSD, in the spirit of the ds4 local inference engine for Metal and CUDA.[1][18] The core proposition is that many developers and power users want high‑quality, open‑weight models with strong reasoning, coding, and multimodal capabilities, but are blocked by memory limits, context constraints, and complex tooling on their personal machines.[2][6][13][18] This report investigates organic demand signals for that idea across six channels—Reddit, Hacker News, Product Hunt, X/Twitter, SEO, and macro trends—using only verifiable examples from 2024–2025 where possible and explicitly flagging any gaps. The evidence shows clear frustration around running large models locally on Apple Silicon, growing enthusiasm for MoE architectures that decouple total parameter count from active compute, emerging interest in SSD offloading and partial loading, and multiple adjacent launches in the local‑LLM tooling space. However, direct Reddit and X/Twitter examples matching the precise combination of “Qwen‑class MoE on 16 GB M1 Pro with SSD streaming” could not be verified in the provided sources, and reliable keyword volume data is absent, so some aspects of demand must be inferred from adjacent signals rather than quantified directly.[2][3][6][8][15][16][18][19] ## Technical And Market Context For The Business Idea ### The ds4 Engine, Qwen MoE Models, And Apple Silicon Constraints The starting point for understanding organic demand is to clarify what the proposed “Show HN” product actually does and why it is differentiated. The GitHub profile for the developer behind ds4, andreaborio, describes ds4 as a “Flash local inference engine for Metal and CUDA,” indicating that it is a performance‑oriented runtime for local large language model inference on GPUs, including Apple’s Metal backend.[1] A “Flash” inference engine in this context typically refers to optimizations analogous to FlashAttention and other memory‑efficient attention implementations designed to maximize throughput and context length given limited memory bandwidth and capacity, which is exactly the constraint faced by users of 16 GB M1 Pro machines.[6][14] The mention of Metal and CUDA suggests cross‑platform support, but the business idea here focuses explicitly on Apple Silicon, where unified memory and SSD performance make certain offloading strategies more attractive than on traditional discrete‑GPU systems.[6][8][19] The model family targeted by the idea, Qwen3.x‑35B‑A3B, is part of Alibaba’s line of open‑weight foundation models; Qwen3.5‑35B‑A3B is documented on Hugging Face as a vision‑language‑capable model with a hybrid architecture combining Gated Delta Networks and sparse Mixture‑of‑Experts routing.[13] This design yields a unified multimodal foundation that reportedly matches or exceeds prior Qwen3 and Qwen3‑VL models in reasoning, coding, agentic behavior, and visual understanding benchmarks, while maintaining efficiency.[13] A later overview from Labellerr describes Qwen3.6‑35B‑A3B as a hybrid MoE model with 35 billion total parameters but only roughly 3 billion parameters active per token, emphasizing that this design allows large total capacity while keeping per‑token computation and memory footprint low enough for constrained hardware.[18] From a business perspective, this “35B total, 3B active” structure is crucial, because it means that with careful routing and offloading, users with modest RAM can access quality more traditionally associated with much larger dense models.[15][18] Apple Silicon machines introduce a distinctive constraint profile that shapes demand. A 16 GB M1 Pro MacBook Pro has a unified memory pool shared between CPU, GPU, and Neural Engine rather than separate VRAM, which removes discrete GPU VRAM ceilings but replaces them with a single tighter system‑wide cap.[6][8] SitePoint’s 2026 guide to local LLMs on Apple Silicon notes that the unified memory architecture of M1, M2, and M3 chips eliminates the biggest constraint that historically limited consumer PCs for large‑model inference—GPU VRAM—but simultaneously warns that the operating system and background processes still consume a substantial fraction of that unified pool.[6] The same guide gives a
| Stage | Total/mo | Breakdown |
|---|---|---|
| M1 (~10) | $25 | Website Hosting (Vercel) $0 + Cloudflare R2 Storage (Base models) $10 + Resend $15 |
| M6 (~100) | $45 | Website Hosting $0 + Cloudflare R2 Storage/Ops $30 + Resend $15 |
| M12 (~1K) | $125 | Website Hosting $20 (Pro) + Cloudflare R2 $70 + Sentry $35 |
| Month | MRR |
|---|---|
| M1 | $0 |
| M3 | $0 |
| M6 | $150 |
| M12 ✅ Breakeven | $400 |
| Month | P20 | P50 realistic | P80 |
|---|---|---|---|
| M1 | $0 | $0 | $200 |
| M3 | $0 | $300 | $1500 |
| M6 | $150 | $150 | $5000 |
| M12 | $400 | $400 | $14000 |