Qwen on Mac (2026): Best Model for 16GB, 24GB, 32GB RAM

TL;DR: On a 24GB Mac, run Qwen 3.8 27B (ollama run qwen3.8:27b), Qwen's dense August 2026 model with vision. Qwen's model card reports 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro. It is a tight fit at Q4, so keep the context short and other apps closed. At 32GB, the Qwen 3.6 35B-A3B MoE (only 3B active per token) is the stronger all-rounder. At 16GB, Qwen 3.5 9B is the largest Qwen that fits. Pick by RAM tier below.

Qwen 3.8, 3.6 or 3.5: Which Should You Run on Mac?

Short answer: pick by unified memory. ModelFit's engine budgets about 70% of RAM for model weights on Macs up to 32GB, because macOS, your apps and the context window need the rest. On a 24GB Mac that leaves roughly 17GB, so the 27B models sit right at the edge and the 35B-A3B MoE models start at 32GB.

The newest model has its own install guide: Run Qwen3.8-27B Locally. For the Qwen 3.6 RAM-by-chip table (M1-M5) and MLX speeds, see the dedicated Qwen 3.6 on Mac guide. The rest of this page details the Qwen 3.5 Medium series, the release that made the 35B-A3B MoE design popular.

Qwen 3.6 on MacBook Pro and M5 Max: Which Size Fits?

Qwen 3.6 ships in two base sizes for local Macs, each with a standard and a Q8 quantization, four Ollama tags total. All four are registry-verified. Pick by how much unified memory your MacBook Pro has.

On a 24GB MacBook Pro (base M4 or M5), neither Qwen 3.6 Q4 tag fits ModelFit's memory budget: both start at 32GB. Run Qwen 3.8 27B instead (ollama run qwen3.8:27b) and keep the context window short.

On a 32GB MacBook Pro, either Q4 tag fits: ollama run qwen3.6:27b for the dense coder, or ollama run qwen3.6:35b-a3b if you want the MoE model's faster responses.

On a 48GB MacBook Pro (M4 Pro, M5 Pro), step up to the Q8 27B (ollama run qwen3.6:27b-q8_0) for near-lossless quality at the same dense-model reasoning depth.

On an M5 Max MacBook Pro, unified memory scales up to 128GB, so every Qwen 3.6 tag in the table above fits comfortably, including the 64GB-minimum 35B-A3B (Q8). That is the quality ceiling for this lineup on a Mac today: ollama run qwen3.6:35b-a3b-q8_0.

Estimated tok/s by chip and quantization, plus the full M1 through M5 breakdown, live on the dedicated Qwen 3.6 on Mac guide.

Alibaba's Qwen 3.5 Medium series remains a masterclass in one thing: smart architecture beats raw parameters. And in April 2026, the Qwen ecosystem has grown even stronger.

Memory each Qwen 3.5 Medium model needs at Q4_K_M, from the tables below. The Qwen team's own benchmark chart is on the Qwen3.5-35B-A3B model card.

April 2026 Update: Since this article was first published, Alibaba shipped Qwen 3.6-Plus (API-only flagship), Qwen3-Coder-Next (80B MoE coding model you can run locally), four new Qwen 3.5 Small models (0.8B-9B), and Ollama gained MLX support, which roughly doubled decode speed in Ollama's own test on Apple Silicon. Scroll to the April 2026 Update section for full details.

What Are the Qwen 3.5 Medium Models?

Why Is Qwen3.5-35B-A3B Special?

This model is special.

MoE Architecture: 35B total parameters, only 3B active per token

• 35B total, but only 3B active thanks to MoE (Mixture of Experts) architecture
• It edges out Qwen3-235B-A22B on MMLU-Pro (85.3 vs 84.4 in the model card), a model 7× larger
• It runs on a 32GB MacBook Pro (about 20GB at Q4, more than a 24GB Mac can spare)

Translation: You get near-frontier quality on a standard laptop, without needing a Mac Studio 512GB.

Why Is This Possible?

Alibaba focused on:

• Better architecture (smarter expert routing)
• Data quality (less but better)
• Optimized RL (Reinforcement Learning)

The result: More intelligence with less compute.

How Does It Compare on Benchmarks?

Qwen3.5-35B-A3B vs Qwen3-235B-A22B

Scores from the Qwen3.5-35B-A3B model card.

💡 The model 7× smaller matches the 235B on knowledge benchmarks and reads about 7× fewer active weights per token, so it decodes faster on the same Mac.

Which Mac?

Speeds are ModelFit estimates at Q4, fitted on measured runs, not ModelFit benchmarks, and capped by each chip's memory bandwidth. Typical error is about 10%. A 48GB M4 Pro roughly doubles the 27B and the 35B-A3B (about 14 and 85 tok/s est.). Qwen3.5-Flash is API-only, so it has no Mac row.

April 2026 Update: What's Changed

The Qwen ecosystem has expanded significantly since February. Here is everything that matters for Mac users running models locally.

Qwen 3.6-Plus: New API Flagship (April 2, 2026)

Alibaba released Qwen 3.6-Plus on April 2, 2026. It is a closed-source model available only via API (OpenRouter, Alibaba Cloud), with no weights and no parameter count to size against a Mac.

What this means for local users: Qwen 3.6-Plus is API-only, so you cannot run it via Ollama. The open-weight Qwen 3.5 Medium series remains your best option for local inference. However, 3.6-Plus shows where the architecture is heading. Expect open-weight successors later in 2026.

Qwen3-Coder-Next: Best Local Coding Model (March 2026)

Qwen3-Coder-Next is a specialized 80B MoE model with only 3B active parameters, designed specifically for agentic coding.

Mac requirement: the Q4 build is 52 GB on Ollama, more than ModelFit budgets for weights on a 64GB Mac, so plan on 96GB or more. For MacBook Pro 24-32GB, the 35B-A3B remains the better all-rounder.

# Run Qwen3-Coder-Next locally
ollama run qwen3-coder-next

Qwen 3.5 Small Models: Tiny but Powerful (March 2026)

Alibaba released four small models on March 2, 2026: 0.8B, 2B, 4B, and 9B. All are natively multimodal (text + image + video), Apache 2.0 licensed, and support 256K context.

The 9B model is remarkable: it edges out GPT-OSS-120B (a model 13x its size) on GPQA Diamond (81.7 vs 80.1). For MacBook Air users with 8GB RAM, the 4B is now a strong recommendation.

# Small models: run on any Mac
ollama run qwen3.5:9b
ollama run qwen3.5:4b
ollama run qwen3.5:2b
ollama run qwen3.5:0.8b

Ollama 0.19 + MLX: Decode Nearly Doubles on Apple Silicon (March 31, 2026)

This is the biggest quality-of-life update for Mac users. Ollama 0.19 now uses Apple's MLX framework natively, delivering:

• Prefill from 1,154 to 1,810 tokens/s (time to first token)
• Decode from 58 to 112 tokens/s (generation speed)
• Smarter unified memory management
• M5/M5 Pro/M5 Max GPU Neural Accelerator support

Practical impact: those are Ollama's figures for Qwen3.5-35B-A3B on its own Apple silicon test machine (NVFP4 on 0.19 against Q4_K_M on 0.18). Speed still tops out at your chip's memory bandwidth: ModelFit's engine estimates about 39 tok/s for the same model on a 32GB M4 MacBook Pro and about 85 tok/s on a 48GB M4 Pro (est., not measured).

October 2026: MLX is no longer a preview. Since v0.40.0 it is Ollama's default engine on Apple Silicon, see Ollama v0.40.0 runs models on MLX by default.

Updated Qwen Ecosystem Map (April 2026)

Which Model Fits Your Use Case?

Qwen3.5-9B: Best for 8-16GB Macs

• Natively multimodal (text, image, video)
• Matches models 13x its size on reasoning benchmarks
• Ideal for: Vision tasks, general chat, lightweight coding, edge deployment

Qwen3.5-Flash: Production Agents

• 1M context by default (analyze entire documents)
• Built-in tools (calculator, search, code execution)
• Ideal for: Autonomous agents, advanced RAG, long document analysis

Qwen3.5-35B-A3B: The Sweet Spot

• Near-frontier quality on a standard MacBook
• About 85 tok/s on a 48GB M4 Pro MacBook Pro (ModelFit estimate, not measured)
• Ideal for: Coding, complex reasoning, intelligent chatbots

Qwen3-Coder-Next: Dedicated Coding Agent

• 80B MoE with 3B active, 256K context
• 52 GB at Q4, so a 96GB Mac or more
• Ideal for: Agentic coding, repo-level tasks, multi-file refactoring

Qwen3.5-122B-A10B: Local Frontier

• GPT-4 / Claude 3.5 level quality
• Requires Mac Studio Ultra
• Ideal for: Enterprise, research, critical tasks

How Do You Run Them?

Via Ollama (Local)

All Qwen 3.5 models are available on Ollama. Ollama 0.40 and later run them on MLX by default on Apple Silicon.

# Small models (any Mac)
ollama run qwen3.5:9b        # 6.6 GB, best small model
ollama run qwen3.5:4b        # 3.4 GB, great for 8GB Macs

# Medium models (24GB+ Macs)
ollama run qwen3.5:35b-a3b   # 20 GB, best perf/RAM ratio
ollama run qwen3.5:27b       # 16 GB, dense, reliable for coding

# Large models (Mac Studio)
ollama run qwen3.5:122b-a10b # 72 GB, frontier quality
ollama run qwen3-coder-next  # 52 GB, best coding agent

# Tiny models (edge / embedded)
ollama run qwen3.5:2b        # 2.7 GB
ollama run qwen3.5:0.8b      # 1.0 GB

Tip: Ollama 0.40 and later use MLX by default on Apple Silicon. Check with ollama --version and update if needed.

Note: Qwen3.5-Flash is the hosted production version of the 35B-A3B (Qwen model card). There is no Ollama tag for it. Use the API endpoints below for Flash.

Via Cloud API

• Qwen Chat: chat.qwen.ai
• Alibaba Cloud API: modelstudio.console.alibabacloud.com
• Qwen 3.6-Plus (flagship): Available on OpenRouter

ModelFit's Picks by Mac Configuration

Here are our picks by Mac configuration. Speeds are ModelFit estimates fitted on measured runs, not ModelFit benchmarks, with a typical error of about 10%:

MacBook Air M2/M3/M4 8GB

→ Qwen3.5-4B (3.4 GB, ~29 tok/s est. on an M2): Multimodal, surprisingly capable

MacBook Air M3/M4 16GB

→ Qwen3.5-9B (6.6 GB, ~17 to 19 tok/s est. on an M3 or M4): Beats models 13x its size

MacBook Air or Pro 24GB

→ Qwen3.8-27B (ollama run qwen3.8:27b): the newest dense Qwen, a tight fit at Q4, so keep the context short

MacBook Pro M4/M5 32GB

→ Qwen3.6-35B-A3B (about 22 GB at Q4) ⭐ Best all-rounder: the MoE successor of Qwen3.5-35B-A3B, only 3B active per token

Mac Studio 64GB

→ Qwen3.6-35B-A3B at Q8 (ollama run qwen3.6:35b-a3b-q8_0, 64GB minimum): near-lossless build of the 32GB pick

Mac Studio 96GB+

→ Qwen3-Coder-Next (52 GB at Q4): Best local coding agent

→ Qwen3.5-122B-A10B (72 GB, ~69 tok/s est. on an M3 Ultra): Frontier quality

Conclusion

Qwen 3.5 marked a turning point in February 2026, and the ecosystem has only gotten stronger since. The era of "bigger just because" models is over. Well-designed MoE architecture delivers frontier performance with 7x better efficiency.

As of October 2026, the Qwen stack covers every Mac: the 4B for 8GB MacBook Airs, the 9B for 16GB, Qwen 3.8 27B for 24GB, the 35B-A3B MoE from 32GB, and Coder-Next for Mac Studio coding workstations. Ollama's MLX engine, the default since v0.40, nearly doubled decode speed in Ollama's own test, though your chip's memory bandwidth still sets the ceiling.

For modelfit.io users: Qwen 3.8 27B is our pick for a 24GB Mac, and the Qwen 3.6 35B-A3B for 32GB. For dedicated coding, add Qwen3-Coder-Next if you have the RAM. And if you are on a budget Mac, the 9B punches far above its weight.

Related: Compare Qwen 3.5 with DeepSeek-V3 in our head-to-head comparison, browse all Qwen models, the best LLM for MacBook guide ranked by RAM tier, or check MacBook Pro and Mac Studio recommendations.

Frequently Asked Questions

How does Qwen 3.5-35B-A3B beat models 7x its size?

The MoE (Mixture of Experts) architecture activates only 3B of the 35B total parameters per token. Alibaba focused on smarter expert routing, higher quality training data, and optimized reinforcement learning. The result: 85.3 on MMLU-Pro with about 20GB of weights at Q4, against 84.4 for the 235B model that needs about 140GB (model card).

Can I run Qwen 3.5-35B-A3B on a MacBook Air?

Not comfortably. At Q4 the model weighs about 20GB, more than the roughly 17GB ModelFit budgets for weights on a 24GB Mac, so the engine lists it from 32GB. On a 24GB MacBook Air, run Qwen 3.8 27B (ollama run qwen3.8:27b) with a short context instead. On a 32GB MacBook Pro, the 35B-A3B fits. For 16GB MacBook Airs, use the 9B model.

What is the difference between Qwen 3.5 Flash and 35B-A3B?

Flash offers 1M token context (vs 128K for 35B-A3B), making it ideal for analyzing entire documents and codebases. The 35B-A3B delivers higher quality per token for shorter tasks. Both need similar RAM (~20-22GB). Choose Flash for RAG and long documents, 35B-A3B for coding and reasoning.

How fast is Qwen 3.5-35B-A3B on Apple Silicon after the MLX update?

ModelFit's engine estimates about 39 tok/s on a 32GB M4 MacBook Pro and about 85 tok/s on a 48GB M4 Pro (est., not measured), because decode speed is capped by memory bandwidth. Ollama's own MLX test took this model from 58 to 112 tok/s on its Apple silicon test machine (Ollama blog). The MoE architecture keeps speed high because only 3B parameters compute per token.

Is Qwen 3.5-122B worth the extra RAM over 35B-A3B?

The 122B model scores higher on quality benchmarks but needs 72GB RAM and a Mac Studio Ultra. For most users, the quality difference does not justify the 3.5x RAM increase. The 35B-A3B is the practical choice unless you need frontier-level accuracy.

Should I use Qwen3-Coder-Next or 35B-A3B for coding?

Qwen3-Coder-Next is purpose-built for agentic coding: multi-file refactoring, repo-level tasks, and autonomous debugging. However, its Q4 build is 52GB. If you have a Mac with 96GB or more, use Coder-Next for dedicated coding work. On a 32GB MacBook Pro, the 35B-A3B handles coding well alongside general reasoning tasks.

What is Qwen 3.6-Plus and can I run it locally?

Qwen 3.6-Plus is Alibaba's flagship model released April 2, 2026. It is API-only. The weights are not publicly available. For local inference on Mac, the Qwen 3.5 series remains the best open-weight option. Access 3.6-Plus via OpenRouter or Alibaba Cloud.

How does Qwen 3.5-9B compare to larger models?

The 9B model is the standout of the Small series. It edges out GPT-OSS-120B (13x its size) on GPQA Diamond (81.7 vs 80.1), a notable result for a much smaller model, though GPT-OSS-120B stays stronger on harder math benchmarks like HMMT. It fits any Mac with 16GB, where ModelFit's engine estimates about 19 tok/s on an M4 MacBook Air (est., not measured). For users who cannot afford the 32GB the 35B-A3B needs, the 9B is the clear second choice.

Can I run Qwen 3.6 27B on a MacBook Pro?

Yes, from 32GB. The Q4 tag (ollama run qwen3.6:27b) is about 18GB, over ModelFit's budget for a 24GB Mac, so it runs on any MacBook Pro with 32GB or more. On a 24GB MacBook Pro, run Qwen 3.8 27B instead and keep the context short. On a 48GB MacBook Pro, use the Q8 tag (ollama run qwen3.6:27b-q8_0) instead for near-lossless quality. Full RAM-by-chip and MLX speed detail lives on the Qwen 3.6 on Mac guide.

Can an Intel MacBook Pro (i9, 16GB) run Qwen, and how fast is it?

Yes, but slowly. Intel Macs get no Metal GPU acceleration in Ollama: that requires Apple Silicon unified memory, so an i9 MacBook Pro runs models on CPU only. For coding, the right family is Qwen2.5-Coder (often searched as "Qwen 2 coder"): ollama run qwen2.5-coder:7b needs about 4.7GB at Q4 and fits a 16GB machine. Expect single-digit tokens per second on an i9 CPU (ModelFit estimate, not measured), versus roughly 10x that on an M-series Mac. If that feels too slow, drop to ollama run qwen3.5:4b (3.4GB), the newer small model that is much quicker on CPU. For usable local coding speeds, any Apple Silicon Mac with 16GB, even a base M1, outperforms an Intel i9 here.

What Qwen 3.6 model should I run on an M5 Max MacBook Pro?

An M5 Max MacBook Pro scales up to 128GB unified memory, so it comfortably fits every Qwen 3.6 tag, including the Q8 35B-A3B (ollama run qwen3.6:35b-a3b-q8_0, 64GB minimum), the highest-quality quantization in the lineup. For pure coding accuracy, the dense 27B Q8 (ollama run qwen3.6:27b-q8_0) is the alternative top pick.

---

Article originally published February 24, 2026. Last updated April 4, 2026 with Qwen 3.6-Plus, Qwen3-Coder-Next, Qwen 3.5 Small series, and Ollama 0.19 MLX speed improvements. Updated July 2, 2026 with a dedicated Qwen 3.6 MacBook Pro and M5 Max RAM-fit section covering all four registry-verified Ollama tags. Updated July 12, 2026 with an Intel MacBook Pro (i9) answer. Updated October 12, 2026: Qwen 3.8 27B is the 24GB pick, and every RAM floor on this page now matches ModelFit's engine (the Qwen 3.6 Q4 tags and the 35B-A3B models start at 32GB). Updated October 13, 2026: every speed is now a ModelFit engine estimate labeled est., the benchmark rows use the model card's MMLU-Pro and MMLU-Redux scores, Ollama's MLX gains are quoted from its own chart, Qwen3-Coder-Next is sized from its 52GB Ollama build, and figures we could not find in their sources were removed. Updated October 20, 2026: the speed estimates follow ModelFit's new performance model. Updated October 22, 2026: every speed on this page re-checked against the engine as it runs today. Resources:

• HuggingFace Qwen3.5
• Qwen3.5-9B model card
• Qwen3.5-122B-A10B model card
• Ollama Qwen3.5 Library
• Ollama Qwen3-Coder-Next
• Ollama MLX Announcement
• Qwen Chat
• Qwen 3.6-Plus on OpenRouter
• Benchmark: Local vs Cloud
• All Qwen Models: Variants, RAM & Benchmarks
• DeepSeek Models
• Llama Models