There Is No DeepSeek R2 — The Real DeepSeek Path for Macs

What Happened to "DeepSeek R2"?

After R1's landmark release in January 2025, an R2 successor was widely rumored, widely speculated about, and — in some aggregator datasets — listed as if it had shipped. It never did. DeepSeek published no R2 weights, no R2 model card, and no R2 announcement. The company moved its reasoning work into the main V-series instead, and the R-branded line simply ended at R1.

An earlier version of this page treated R2 as a released model, based on data we could not ultimately trace to any official source. The LLMCheck catalog has since been reconciled against vendor releases; every DeepSeek entry in the leaderboard now maps to weights you can actually download.

The Real DeepSeek Line: R1 → V3.2 → V4

V4-Flash: The Actual Mac Story

DeepSeek V4-Flash is a 284-billion-parameter MoE with roughly 13 billion active per token, released in July 2026 under a plain MIT license. The small active set is what makes Apple Silicon speeds respectable despite the huge total footprint; the total parameter count is what makes the memory bill so steep.

Capability-wise it sits in the top open tier: an Artificial Analysis Intelligence Index score of 50, and a self-reported 82.7 on Terminal-Bench 2.1 — a strong agentic-coding result, though it is DeepSeek's own number rather than an independent harness. Either way, this is genuine frontier-class capability under a license with no strings attached.

The headline for this site: the 2-bit MLX build weighs 96.5GB, which means a 128GB Mac can hold the whole model in Unified Memory. That is the first time a member of DeepSeek's flagship generation has fit on Apple Silicon at usable speed.

Mac Hardware Requirements

• 16GB / 24GB / 36GB / 64GB Macs — Not possible. Even the most aggressive quantization of a 284B model is far out of reach.
• 96GB Macs — No. The 96.5GB build cannot coexist with macOS and your apps.
• 128GB Macs (M3 Max, M4 Max, M5 Max, Mac Studio configs) — Yes, via the 2-bit MLX build. Headroom is tight; close memory-heavy apps before loading.

One honest caveat carries over from every giant-model-on-a-Mac story: 2-bit is aggressive quantization. Community reports describe the quantized model as strong, but expect some degradation versus the full-precision weights served via API. As always, every figure in the LLMCheck index carries a provenance label — see the methodology.

Speeds on Apple Silicon

According to the LLMCheck index, the community-reported numbers for the 2-bit MLX build — running on antirez's ds4 Metal engine — are:

These are interactive-grade speeds for a model of this class — a different world from the single-digit tok/s that giant dense models manage on Macs. Long reasoning traces are still time-consuming, but ~39 tok/s makes deliberate agentic and coding work genuinely practical. Full per-chip detail lives on the V4-Flash on M5 Max page and in the benchmarks table.

How to Run It Today

1. MLX via ds4 — antirez's ds4 Metal engine is where the community speed figures above come from, and is currently the most-proven path for the 2-bit MLX build on 128GB Macs.
2. llama.cpp — merged DSpark speculative decoding for the V4 family in August 2026, which meaningfully improves throughput on supported builds.
3. Ollama — not yet supported as of August 2026. If you live in Ollama, watch its release notes; until then, the two paths above are the way in.

The weights themselves are on Hugging Face under MIT. Budget roughly 100GB of disk for the 2-bit build and expect a long first download. See the guides hub for memory-tuning tips on large MoE models.

What About V4 Pro?

DeepSeek V4 Pro — the 1.6T-total / 49B-active flagship — reached general availability in August 2026. Two things to know: it is server-class and will never fit on a Mac at any quantization, and the GA build's weights were not published at launch. The April preview weights remain the open MIT release; the GA model is a hosted product for now. For Mac purposes, V4-Flash is the DeepSeek you run; V4 Pro is the DeepSeek you rent.

Smaller-Mac Alternatives

Most Macs are not 128GB Macs. If yours is smaller, the reasoning-and-coding picks from the LLMCheck index are:

• Qwen3.6-27B — the current LLMCheck Mac #1 (Score 72). A dense 27B under Apache 2.0 scoring 77.2% on SWE-bench Verified, at an estimated ~40 tok/s on an M5 Max. Fits 24GB+ Macs at Q4.
• GLM-4.5-Air — for 64GB Macs. A 106B-A12B MoE under MIT that the index puts at roughly 30 tok/s on a 64GB machine — the biggest-model feel you can get below the 128GB tier.

The practical rule: own a 128GB Mac and want the biggest open DeepSeek that fits — run V4-Flash. Own anything smaller — run Qwen3.6-27B and lose remarkably little.