# m5 max Published articles for m5 max. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Junie Local became 70% faster on Mac M5 DevFeed: [Junie Local became 70% faster on Mac M5]() Original publisher: [Read original article]() Author: Stanislav Erokhin Published: 2026-10-28T14:40:49Z Content type: release Language: en Sources: [JetBrains]() Topics: [junie](), [Local AI](), [LLM optimization](), [quantization](), [AI Foundation Models]() Tags: [acceptance](), [acceptance-rate](), [agent](), [agentic](), [agentic-ai](), [agentic-coding](), [ai](), [attention](), [coding](), [coding-tasks](), [draft](), [engine](), [fast](), [faster](), [inference](), [inference-engine](), [jetbrains-ai](), [junie](), [local](), [local-ai](), [m5-max](), [mac](), [tokens]() ### AI overview Junie Local reports roughly 70 tokens/s on M5 Max, up from about 40, with bursts to 107; prefill rises from 850 to 1,200 tokens/s, and a 100k context uses about 24 GB of RAM. The gains come from replacing MLX-VLM with Splash, which verifies eight drafted tokens per decode step and reduces context attention cost, plus a DFlash2 drafter fine-tuned on Junie Local coding trajectories for Qwen3.8-3.6-27B-blend. The drafter's acceptance rate depends on how familiar a task is, so speed can vary. The release supports M4 Macs and newer with 64 GB RAM, but M4 prefill is about four times slower than M5. Upgrade with `junie update`, then run `/local` and press `ctrl + r` to update the inference engine. ### Source excerpt Junie Local now delivers ~70 tokens/s on M5 Max for coding tasks, up from ~40 tokens/s, with bursts up to 107 tokens/s. Two changes made this possible: the Splash inference engine and a new DFlash2 draft model. Back in August, when we first released Junie Local for Mac M5, we spent a lot of time [...] ## Avrea is the fastest ARM runner for GitHub Actions: 2026 benchmark DevFeed: [Avrea is the fastest ARM runner for GitHub Actions: 2026 benchmark]() Original publisher: [Read original article]() Author: Tatu Mäkijärvi Published: 2026-10-23T00:00:00Z Content type: news Language: en Sources: [Avrea Blog]() Topics: [Benchmark](), [GitHub Actions](), [self-hosted runner](), [CI/CD](), [GitHub]() Tags: [2026](), [apple](), [apple-m5](), [arm](), [arm64](), [benchmark](), [benchmarks](), [build](), [cargo](), [ci](), [fast](), [faster](), [free-tier](), [github](), [github-actions](), [m5-max](), [runner]() ### AI overview In RunsOn's November 2026 cold Rust build benchmark, Avrea's Apple M5 Max ARM runners finished first among 36 runners from seven providers at 2, 4 and 8 vCPU, taking 116, 76 and 58 seconds. GitHub's ARM runners took 316 and 196 seconds at 2 and 4 vCPU; Avrea was 2.7x and 2.6x as fast at those sizes. The benchmark used median build-step times from five cold runs of spotify-player with no cache. RunsOn reports higher arm64 single-thread PassMark scores for Avrea's M5 Max, a factor it links to Rust's serial compilation path. Estimated full-job cost at 2 vCPU was $0.025 on Avrea versus $0.030 on GitHub after GitHub's whole-minute rounding, before allowances; GitHub-hosted runners are free for public repositories. To use Avrea ARM runners, install its GitHub App and choose a runner label. Ubuntu 24.04 and 26.04 sizes range from 1 to 16 vCPU; the free tier covers about 500 minutes on the 4-vCPU, 16-GB runner, shared with other usage. ### Source excerpt RunsOn's public benchmark of 36 arm64 GitHub Actions runners from seven providers: Avrea's M5 Max runners finished first at 2, 4 and 8 vCPU. ## Faster Gemma 4 on MLX with multi-token prediction DevFeed: [Faster Gemma 4 on MLX with multi-token prediction]() Original publisher: [Read original article]() Published: 2026-06-29T00:00:00Z Content type: release Language: en Sources: [Ollama]() Topics: [gemma](), [MLX](), [agentic-coding](), [Ollama](), [Language models](), [AI Infrastructure]() Tags: [2026](), [agents](), [apple-silicon](), [benchmark](), [coding](), [coding-agents](), [gemma](), [gpu](), [hardware](), [m5-max](), [macos](), [mlx](), [model](), [models](), [mtp](), [ollama](), [performance](), [silicon](), [token]() ### AI overview Ollama 0.31 adds multi-token prediction for Gemma 4 on Apple Silicon, with the article reporting nearly 90% faster token generation on average across a coding-agent benchmark. The feature dynamically adjusts draft length, verifies proposed tokens in batches, and uses an MLX kernel to reduce repeated weight reads. The speedup depends on workload and hardware. ### Source excerpt Gemma 4 is now significantly faster in Ollama 0.31 on Apple Silicon via multi-token prediction (MTP), powered by MLX. Performance is now up to 90% faster when used with coding agents, as measured using the Aider polyglot benchmark. ## M5 Max runners are here: macOS and ARM Linux CI on Apple Silicon DevFeed: [M5 Max runners are here: macOS and ARM Linux CI on Apple Silicon]() Original publisher: [Read original article]() Author: Hannu Valtonen Published: 2026-06-09T00:00:00Z Content type: release Language: en Sources: [Avrea Blog]() Topics: [macOS](), [Linux](), [Xcode]() Tags: [actions](), [apple](), [apple-silicon](), [arm](), [arm64](), [build](), [build-system](), [builds](), [built-in](), [c](), [c-plus-plus](), [cache](), [caching](), [ci](), [github](), [github-actions](), [linux](), [m5-max](), [macos](), [performance]() ### AI overview Avrea offers GitHub Actions runners on M5 Max Apple Silicon for macOS and ARM Linux, with caching intended to reduce repeated dependency downloads and builds. The article says its 8-core macOS runner costs $0.08 per minute, compared with GitHub's M2 Pro 5-core runner at $0.102 per minute (figures as of writing). It describes Xcode caching for Apple-platform workflows and ecosystem caches for ARM Linux, including Go, Rust, Node.js, Python, C/C++ and Zig when configured. Teams can generally try the runners by changing the workflow's `runs-on` label; the article recommends matching runner size to observed CPU and memory use. It names example labels for macOS and Ubuntu ARM jobs, while directing readers to current labels for available sizes. ### Source excerpt Avrea now supports GitHub Actions workloads on macOS and ARM Linux powered by the latest M5 Max Apple Silicon. Less money for much more performance. ## Ollama is now powered by MLX on Apple Silicon in preview DevFeed: [Ollama is now powered by MLX on Apple Silicon in preview]() Original publisher: [Read original article]() Published: 2026-03-30T00:00:00Z Content type: release Language: en Sources: [Ollama]() Topics: [MLX](), [Ollama](), [AI tools for developers](), [Local AI](), [Language models](), [GPU]() Tags: [apple](), [apple-silicon](), [claude](), [claude-code](), [codex](), [coding-agents](), [cuda](), [ggml](), [gpu](), [inference](), [llama-cpp](), [m5-max](), [m5-pro](), [memory](), [mlx](), [ollama]() ### AI overview Ollama previews a version powered by Apple's MLX framework for Apple silicon, with reported performance gains and updated cache reuse and checkpointing for coding and agentic tasks. It also adds NVFP4 model support and provides commands to try a Qwen3.5 coding model on a Mac with more than 32 GB of unified memory. ### Source excerpt Today, we're previewing the fastest way to run Ollama on Apple silicon, powered by MLX, Apple's machine learning framework.