LLMCheck BlogLocal LLMs on Apple Silicon Macs: model reviews, hardware guides and the monthly State of Open-Source Local LLMs edition.https://llmcheck.net/blog/2026-10-18T00:00:00ZLLMCheckhttps://llmcheck.nethttps://llmcheck.net/favicon-32x32.pngNemotron 3.5 Bolt on a Mac: The Metal Crash, What Works, Super VLhttps://llmcheck.net/blog/nemotron-3-5-bolt-super-vl-mac/2026-10-18T00:00:00Z2026-10-18T00:00:00ZNemotron 3.5 Bolt segfaults in stock llama.cpp on Metal and Ollama marks it unsupported on Apple Silicon. What runs today, speeds, and Super VL for 128 GB.Muse Spark Open on a Mac: What Runs on One 512 GB Mac Studiohttps://llmcheck.net/blog/muse-spark-open-mac/2026-10-18T00:00:00Z2026-10-18T00:00:00ZMeta's 816B Muse Spark Open fits one 512 GB Mac Studio at 2–3-bit: 21 tok/s (oQ3e) and 18 tok/s (Q2_K) published on M3 Ultra. 4-bit needs two Macs.How Accurate Are LLMCheck's Mac Speed Estimates? 67 Runs Checkedhttps://llmcheck.net/blog/mac-llm-speed-estimates-vs-measurements-october-2026/2026-10-18T00:00:00Z2026-10-18T00:00:00ZLLMCheck's tok/s formula vs 67 new Mac runs: dense models within 18%, small MoEs right on average, large MoEs at 0.4×–2.1× the estimate.Best Local LLM for a 24GB or 32GB Mac (October 2026)https://llmcheck.net/blog/best-local-llm-24gb-mac-october-2026/2026-10-18T00:00:00Z2026-10-18T00:00:00ZAccording to the LLMCheck index, Qwen 3.6-35B-A3B is the best local LLM for 24GB and 32GB Macs. North Mini Code, KAT-Coder, LFM2.5 and MiniCPM-V by job.Best Local LLM for a 16GB Mac (October 2026)https://llmcheck.net/blog/best-local-llm-16gb-mac/2026-10-18T00:00:00Z2026-10-18T00:00:00ZAccording to the LLMCheck index, Gemma 4 12B is the best local LLM for a 16GB Mac. Picks for coding, speed and ternary 27B models, with file sizes.M5 Ultra vs M3 Ultra vs M5 Max for Local LLMs: The Measured Numbershttps://llmcheck.net/blog/m5-ultra-vs-m3-ultra-vs-m5-max-local-llm/2026-10-18T00:00:00Z2026-10-18T00:00:00ZPublished runs: the M5 Ultra decodes ~1.5× faster than an M3 Ultra and reads prompts 2–4× faster. When an M5 Max or M3 Ultra wins, and what fits 512 GB.Best Local Vision Model for Mac (2026), by RAM Tierhttps://llmcheck.net/blog/best-local-vision-model-mac/2026-10-18T00:00:00Z2026-10-18T00:00:00ZMiniCPM-V 4.7 is the best local vision model for a 24–32 GB Mac. Picks from 8 GB to 512 GB, with memory needs, speeds and llama.cpp, MLX and Ollama setup.State of Open-Source Local LLMs — October 2026: Small MoEs Take the Tophttps://llmcheck.net/blog/state-of-open-source-local-llms-october-2026/2026-10-05T00:00:00Z2026-10-18T00:00:00ZOctober 2026 report (5 Oct, updated 18 Oct): first Mac mini M6 runs, the MoE refit, MiMo-V2.6-Flash to #1, Muse Spark Open, and Ollama goes MLX-first.M5 Ultra and M6, Measured: How the LLMCheck Estimates Held Up (2026)https://llmcheck.net/blog/m5-ultra-m6-measured/2026-09-23T00:00:00Z2026-10-18T00:00:00ZThe first M5 Ultra and Mac mini M6 measurements set against the LLMCheck estimates: BGR, Apple, Tom's Hardware, tbreak and community runs, each labelled.DeepSeek V4.1 Flash and MiMo-V2.6: Two MIT Frontier Models on a Mac (2026)https://llmcheck.net/blog/deepseek-v4-1-flash-mimo-v2-6-mac/2026-09-23T00:00:00Z2026-10-18T00:00:00ZDeepSeek V4.1 Flash (552B, MIT) now runs on a 512 GB Mac Studio through oMLX; MiMo-V2.6-Flash (309B-A15B, MIT, omnimodal) runs from a 192 GB Mac Studio.State of Open-Source Local LLMs — September 2026: Two Frontier MoEs Reach the Mac Studiohttps://llmcheck.net/blog/state-of-open-source-local-llms-september-2026/2026-09-07T00:00:00Z2026-10-05T00:00:00ZSeptember 2026 open-source LLM report: Qwen3.8-27B takes #1, GLM-5.3-Flash (MIT, the Ox Alpha model) and Qwen3.8-Flash-Next land with Mac Studio-class quants, DeepSeek opens V4 Flash Vision, and the index recalibrates its MoE speed estimates.GLM-5.3-Flash on a Mac: Which Studio Runs the Ox Alpha Model (2026)https://llmcheck.net/blog/glm-5-3-flash-on-mac/2026-09-07T00:00:00Z2026-10-18T00:00:00ZGLM-5.3-Flash (320B-A18B, MIT) on a Mac: which quant fits 128 vs 256 GB, tok/s per chip, and running it in stock llama.cpp, oMLX or MLX.Qwen3.8-Flash-Next on a Mac: 6B Active, 112 GB, and the License Clause to Read First (2026)https://llmcheck.net/blog/qwen3-8-flash-next-on-mac/2026-09-07T00:00:00Z2026-10-18T00:00:00ZQwen3.8-Flash-Next, the Qwen 4 architecture preview, needs a 128 GB Mac: 44 tok/s community-measured on an M5 Max. RAM, setup and the license clause.Apple’s New Mac mini and Mac Studio: What M6 and M5 Ultra Mean for Local LLMshttps://llmcheck.net/blog/new-mac-mini-mac-studio-m6-m5-ultra-local-llm/2026-08-25T00:00:00Z2026-10-18T00:00:00ZApple's M6 Mac mini and M5 Max and M5 Ultra Mac Studio for local LLMs: bandwidth, memory, prices and speeds from the LLMCheck index, with provenance.M5 Ultra Mac Studio: 1.2 TB/s and 512 GB, Examined for Local LLMshttps://llmcheck.net/blog/m5-ultra-mac-studio-local-llm/2026-08-25T00:00:00Z2026-10-18T00:00:00ZApple's M5 Ultra for local LLMs: 1.2 TB/s, 50% above the M3 Ultra, and up to 512 GB. Review figures, index speeds and why 256 GB is the config to buy.The $899 M6 Mac mini as a Local-AI Machine: What Actually Fitshttps://llmcheck.net/blog/m6-mac-mini-local-llm/2026-08-25T00:00:00Z2026-10-18T00:00:00ZMac mini M6 for local LLMs: Qwen 3 30B-A3B at 62 tok/s (published), Qwen 3.6-35B-A3B at 64 (community), a dense 27B near 9. Which config to buy.UI-Mate on Mac: RAM, Speed and Setup for Tencent’s Open Computer-Use Agenthttps://llmcheck.net/blog/ui-mate-computer-use-agent-mac/2026-08-21T00:00:00Z2026-08-21T00:00:00ZUI-Mate-27B is Apache 2.0 and runs on Apple Silicon — but its 4-bit build is 17.5 GB and it holds five screenshots in context, so a 24 GB Mac is not enough. Real quant sizes, per-chip speed estimates, and which Mac actually works.State of Open-Source Local LLMs — August 2026https://llmcheck.net/blog/state-of-open-source-local-llms-august-2026/2026-08-13T00:00:00Z2026-08-13T00:00:00ZAugust 2026 open-source LLM report: Meta returns with Apache-2.0 Muse Glimmer 30B, DeepSeek V4 Flash runs on 128GB Macs, Qwen open-weights its 2.4T Max, and the LLMCheck index completes a full provenance audit. Verified Mac #1: Qwen 3.6-27B.GLM 5.2: The Open Model That Beat GPT-5 and Claude — on Its Own Harnesshttps://llmcheck.net/blog/glm-5-2-open-model-beats-gpt5-claude/2026-07-11T00:00:00Z2026-07-11T00:00:00ZZhipu AI's GLM 5.2 (June 2026, MIT, 753B-A40B MoE) reports 68.5% on SWE-Bench Pro — on its own scaffold. On Scale's standardized SEAL harness it scores ~62%, the top open-weights result. Harness-annotated benchmarks, plus the community quants that run it on 256–512 GB M3 Ultra Macs.Apple Silicon Memory Bandwidth & LLM Speed (2026): M1–M5, Pro, Max, Ultrahttps://llmcheck.net/blog/apple-silicon-memory-bandwidth-llm/2026-07-11T00:00:00Z2026-07-11T00:00:00ZWhy memory bandwidth is the #1 factor in local LLM tokens-per-second on a Mac. Full GB/s table for every Apple Silicon chip (M1–M6, Pro/Max/Ultra) and how to estimate tok/s from bandwidth and model size.Qwen 4 Release Date: What Alibaba Announced, and What to Run on a Machttps://llmcheck.net/blog/qwen-4-release-whats-new-mac/2026-07-11T00:00:00Z2026-10-18T00:00:00ZQwen 4 is in training: announced 22 Sep 2026 with no weights, licence or date as of 18 Oct. What Alibaba confirmed, and which Qwen to run on a Mac now.M4 vs M5 for Local LLMs: Is the New Apple Silicon Worth It?https://llmcheck.net/blog/m4-vs-m5-local-llm/2026-07-11T00:00:00Z2026-07-11T00:00:00ZM4 vs M5 Apple Silicon for local AI — what the M5's Neural Accelerators and higher bandwidth actually mean for tokens-per-second. Real benchmark deltas across M4/M5, Pro, and Max, and whether to upgrade.State of Open-Source Local LLMs — July 2026 (Corrected Edition)https://llmcheck.net/blog/state-of-open-source-local-llms-july-2026/2026-07-11T00:00:00Z2026-07-11T00:00:00ZCorrected July 2026 open-source LLM report: GLM 5.2's harness-checked frontier lead, Thinking Machines' Inkling debut, Kimi K3's record 2.8T open weights, DeepSeek V4-Flash on 128 GB Macs, and the on-device ternary wave. Apple Silicon guidance with labeled provenance.GLM 5.2 vs DeepSeek V4: The Real Open-Frontier Showdownhttps://llmcheck.net/blog/glm-5-2-vs-deepseek-r3/2026-07-11T00:00:00Z2026-07-11T00:00:00ZGLM 5.2 (753B, MIT) vs DeepSeek V4-Flash and V4 Pro — harness-annotated scores, license reality, and which open-frontier giant a Mac can actually run. Corrected: 'DeepSeek R3' never existed; the real DeepSeek line is R1 → V3.2 → V4.GLM-4.5-Air on a Mac — and What Happened to "GLM 5.2 Air"https://llmcheck.net/blog/glm-5-2-air-on-mac/2026-07-11T00:00:00Z2026-07-11T00:00:00ZNo 'GLM 5.2 Air' exists — the real Mac-runnable GLM with those specs is GLM-4.5-Air: 106B-A12B MoE, MIT licensed, ~30 tok/s (estimated) on a 64 GB Mac. Corrected setup guide, hardware requirements, and the newer GLM-4.7-Flash option.GLM 5.2 vs Qwen3.8-Max vs Kimi K3: The Open Frontier Three-Way (August 2026)https://llmcheck.net/blog/glm-5-2-vs-qwen-4-1-vs-llama-5-405b/2026-07-11T00:00:00Z2026-07-11T00:00:00ZGLM 5.2 (MIT), Qwen3.8-2.4T and Kimi K3 compared on licence, size and coding scores, and what a Mac can run instead, per the LLMCheck index.Qwen 4 vs Qwen 4.1: What Exists, and Which Qwen to Run on a Machttps://llmcheck.net/blog/qwen-4-1-vs-qwen-4/2026-07-11T00:00:00Z2026-10-18T00:00:00ZQwen 4 was announced 22 Sep 2026 and is still in training, with no open weights; Qwen 4.1 does not exist. What shipped instead, and what to run on a Mac.Phi-5 Doesn't Exist: Phi-4 Is Still Microsoft's Local Line (and What Beats It)https://llmcheck.net/blog/phi-5-large-28b-review/2026-07-11T00:00:00Z2026-07-11T00:00:00ZThere is no Phi-5: Microsoft's local line tops out at Phi-4. Where Phi-4 14B still fits, and what the LLMCheck index picks for 24–32 GB Macs instead.Llama 5 Still Isn't Out — Meta's Real 2026 Release Is Muse Glimmer 30B (August 2026)https://llmcheck.net/blog/llama-5-405b-review/2026-07-11T00:00:00Z2026-10-18T00:00:00ZLlama 5 has not been released. Meta's open model for 24–32 GB Macs is Muse Glimmer 30B: Apache 2.0, image input, 27 tok/s vendor-reported on an M5 Max.State of Open-Source Local LLMs — June 2026https://llmcheck.net/blog/state-of-open-source-local-llms-june-2026/2026-06-06T00:00:00Z2026-06-06T00:00:00ZCorrected June 2026 report, kept as published with notes: its Qwen 4, Llama 5 70B, Voyage Pro, Grok 4 Open and Phi-5 Medium were not real open-weight releases.