\n| Source type<\/td>\n | Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n Values determined by this site’s code at collection time. Dates are JST.<\/em><\/p>\n<\/p>\n Overview<\/h2>\nOn September 18, 2026, Inco AI released “Splash Engine,” an open-source local LLM inference engine tailored specifically for Apple Silicon Mac environments. Alongside this, the company announced a collaboration with LM Studio, natively integrating Splash Engine into the desktop application “LM Studio Bionic” released on the same day. Unlike general-purpose inference engines, Splash Engine adopts a design that builds kernels and memory plans dedicated to specific models, reportedly achieving high processing speeds on supported models.<\/p>\n Announcement Details<\/h2>\nSplash Engine, released under the Apache-2.0 license, is an open-source inference engine available on GitHub (github.com\/incoai\/splash<\/code>). While general inference engines aim for broad compatibility across diverse models, Splash Engine takes an approach of optimizing the entire engine on a per-model basis. At present, it supports two models: “Qwen3.6-35B-A3B” and “Qwen3.8-27B,” featuring model-specific implementations of fused shape-specialized kernels, dedicated trained draft models, and optimized memory plans.<\/p>\nFor inference processing, it incorporates speculative decoding as a standard feature. Each supported model comes bundled with a dedicated draft model named “DFlash 2.” Draft models in the DFlash series also operate in SGLang, vLLM, TensorRT-LLM, and llama.cpp, boasting a track record of over 6 million total downloads. In Splash Engine, token proposal, verification, acceptance, and state updates by the draft model are executed as a single processing unit, reducing per-step overhead. Additionally, it features 4-bit matrix multiplication and attention kernels pre-compiled by internal kernel agents, eliminating the need for manual compilation or tuning by the user.<\/p>\n Memory management involves pre-budgeting according to the hardware configuration. For instance, when running Qwen3.8-27B, 15 GiB is allocated to model weights and 1.2 GiB to the draft model, after which the KV cache area is calculated.<\/p>\n As performance validation, measurement results using “NVIDIA SPEED-Bench” on an M5 Pro Mac environment equipped with 48 GB of unified memory were published. For single-request processing of Qwen3.8-27B, the decode speed recorded 74 tokens\/s for short prompts and 54 tokens\/s for a 32K context, which is reportedly nearly twice the speed of the runner-up “oMLX” among comparison targets. Qwen3.6-35B-A3B achieved 210 tokens\/s for short prompts and 143 tokens\/s at 32K. Furthermore, the Time to First Token (TTFT) when reusing the cache at a 32K context was 123 ms for Qwen3.6-35B-A3B and 282 ms for Qwen3.8-27B, representing a 6.6x to 7.3x speedup compared to oMLX. In addition, validation of parallel requests (short prompts) by four sub-agents reached a total throughput of 170 tokens\/s for Qwen3.8-27B and 357 tokens\/s for Qwen3.6-35B-A3B, showing up to a 3.9x speed improvement over the runner-up engine.<\/p>\n The hardware requirements call for a Mac equipped with an M3 or newer processor, running macOS 26.4 or later, with 36 GB or more of unified memory (Inco AI recommends 48 GB or more). For installation, users can install it via Homebrew using brew install incoai\/tap\/splash<\/code> and launch it with the command splash serve --model incoai\/Qwen3.8-27B-Splash<\/code>, alongside CLI integration commands for tools such as OpenCode, Claude Code, Codex, and Hermes.<\/p>\nAdditionally, through collaboration with LM Studio, integrated use is available in “LM Studio Bionic” (version 1.1.5 or later). By installing Splash (Metal) from Experimental backends under Settings > Runtime in LM Studio Bionic and downloading the supported models from the Hugging Face repository (incoai\/Qwen3.8-27B-Splash<\/code> or incoai\/Qwen3.6-35B-A3B-Splash<\/code>), users can execute local inference directly from the UI.<\/p>\nBackground<\/h2>\nInco AI has previously developed inference platforms based on data center GPUs. On the benchmark evaluation platform “Artificial Analysis,” the company is described as having the record of being the fastest provider across five open models: “Kimi K3,” “MiniMax M3,” “GLM 5.3,” “GLM 5.3 Flash,” and “DeepSeek V4.1 Flash.” Moreover, its proprietary speculative decoding draft model series “DFlash” supports major inference frameworks such as SGLang, vLLM, TensorRT-LLM, and llama.cpp, accumulating expertise in model acceleration with over 6 million total downloads.<\/p>\n In recent years, with the growing demand for local coding assistance and agent operations, high-load processing tasks\u2014such as initial repository loading (prefill), conversation continuation with accumulated history (decode), prefix caching for the same context, and parallel execution by multiple sub-agents\u2014have become routine. However, many conventional local inference engines are designed with a primary focus on the flexibility to execute diverse model structures generally, creating overhead in optimizing for specific model shapes and integrating speculative decoding. Inco AI aims to resolve performance bottlenecks in local agent tasks by applying the philosophy cultivated in data center infrastructure\u2014specializing the entire engine design to match the model\u2014to Apple Silicon Macs.<\/p>\n Impact on Local LLM Users<\/h2>\nFor developers operating open-weight models in local environments, the emergence of Splash Engine provides a high-speed execution option on Apple Silicon environments.<\/p>\n Usage Forms and Accessibility<\/h3>\nSplash Engine is open-sourced under the Apache-2.0 license and can be installed directly as a CLI tool from GitHub via Homebrew. In addition, because it was natively integrated on day one into LM Studio Bionic (1.1.5 or later), a representative local LLM desktop application, it can be deployed and executed from the GUI without complex environment setup. As API interfaces, it supports OpenAI Chat Completions, Responses, and Anthropic Messages formats, and supports streaming output, tool calling, JSON Schema output, and image input. It is designed to connect seamlessly with existing coding agent tools such as OpenCode, Claude Code, Codex, and Hermes.<\/p>\n Hardware Requirements and Constraints<\/h3>\nWhile delivering high performance, the requirements placed on hardware are set relatively high. Supported requirements are specified as a Mac with an M3 or newer processor, macOS 26.4 or later, and at least 36 GB or more of unified memory, with Inco AI recommending 48 GB or more of memory for a practical environment. When running Qwen3.8-27B, in addition to the model weights themselves (15 GiB) and the draft model (1.2 GiB), the KV cache area is reserved, meaning a large amount of unified memory is a prerequisite to process long contexts while running editors and browsers concurrently. It should be noted that older generation M1\/M2 chips and environments with less than 36 GB of memory are out of scope for support.<\/p>\n Supported Models and Future Scalability<\/h3>\nThe “model-specialized” design that is Splash Engine’s core feature comes with the trade-off that choices of supported models are limited. Currently, only two models are provided: “Qwen3.6-35B-A3B” and “Qwen3.8-27B,” and it does not function as a general-purpose runtime that can freely load and execute arbitrary models in general GGUF or MLX formats. Inco AI has expressed its policy to continuously add supported models going forward, stating that it accepts requests for user-customized fine-tuned models and inquiries from agent developers. However, the materials do not mention specific additional model names or timelines outside of the Qwen series.<\/p>\n <\/p>\n |