Best 70B-Class Local Models (That Actually Exist): Llama 3.3 vs Hermes 4 vs Apertus 1.5

TL;DR — Quick Verdict

What Happened to "Llama 5 vs Voyage Pro"

The short version: during a July 2026 data incident, model names that never shipped entered the LLMCheck catalog, and this page compared two of them. The catalog has since been reconciled against the live web, and both entries are gone.

The verified record, as of August 2026: Meta's open flagship is still Llama 4 Maverick, its last dense 70B is Llama 3.3, and the next-generation Llama — reported under the codename "Avocado" — is expected in 2027. Meta's real open release of August 2026 is Muse Glimmer 30B, a 30B multimodal agent model. Mistral's current open line runs through Large 3 (675B-A41B, server-class), Small 4 (119B-A6B MoE), Small 3.2 24B, Ministral 3 14B, and Devstral — there is no dense 70B and no "Voyage" series anywhere in it.

But the question behind the original page is still a good one: what is the best 70B-class model you can actually run on a 64GB+ Mac? Three real candidates stand out.

The Real Field: Head-to-Head Specs

Llama 3.3 70B (Meta, December 2024) is the reference dense 70B — the ecosystem default with the most fine-tunes and the most predictable behavior. Hermes 4 70B (Nous Research, August 2025) is a heavily post-trained Llama 3.1 70B with a hybrid reasoning mode. Apertus 1.5 70B (Swiss AI Initiative — ETH Zürich and EPFL, July 2026) is the transparency outlier: weights, training data, and recipe all published.

Capability sub-scores and speeds are from the LLMCheck index; speeds are estimated at Q4_K_M via MLX. Where vendor benchmark claims are cited below, they are labeled as such — see the methodology page for how estimates are produced.

Estimated Speed on Apple Silicon

Dense 70B inference is memory-bandwidth-bound: every parameter is read on every token, so Ultra-class chips with wider memory buses help more than a newer Max. According to the LLMCheck index (all figures estimated, Q4_K_M):

The index also carries Hermes 4 at ~13 tok/s estimated on an M4 Max, and Llama 3.3 holds up on older silicon — ~15 tok/s estimated on M3 Ultra, ~14 on M1 Ultra, with a practical floor around ~7 tok/s on an M2 Max. None of these numbers is fast by 2026 MoE standards: 11–16 tok/s is comfortable for chat, workable for code review, and slow for long agent loops. Per-chip detail lives on the model pages: Llama 3.3 70B, Hermes 4 70B, Apertus 1.5 70B.

Three Profiles in Detail

Llama 3.3 70B — the proven generalist

Nearly two years old and still the default answer. Meta's vendor-reported card numbers — 86.0 MMLU (0-shot, CoT), 88.4 HumanEval, 77.0 MATH, 92.1 IFEval — hold up remarkably well for a December 2024 model, and its instruction-following reliability is the reason hundreds of fine-tunes use it as a base. It shows its age on agentic benchmarks and has the oldest knowledge cutoff of the three, but nothing in this class is better documented or more widely supported: every Mac runtime ships it out of the box.

Hermes 4 70B — the steerable creative

Nous Research's post-train of Llama 3.1 70B adds a hybrid reasoning mode — the model can emit explicit thinking traces before answering, and you can toggle the behavior off for speed. Its character is the draw: vendor-reported results emphasize markedly lower refusal rates and steerability via system prompt, and the community consensus matches — this is the strongest creative-writing and roleplay model of the trio, with solid math when the reasoning mode is on. It is also, narrowly, the fastest of the three at ~16 tok/s estimated on an M5 Max.

Apertus 1.5 70B — the fully open one

The Swiss AI Initiative's July 2026 update is the only model here that is open source end-to-end: Apache 2.0 weights and published training data and recipe, which makes it the audit-friendly choice for compliance-sensitive deployments. The 1.5 release adds a 262K context window — double its rivals — and multimodal image input, and the model is multilingual by design, trained across 1,000+ languages. The tradeoff is speed: ~11 tok/s estimated on an M5 Max is the slowest of the trio, though its ~20 tok/s estimated on an M4 Ultra makes a Mac Studio the natural host. According to the LLMCheck index it carries the highest capability sub-score of the three at 29/50.

The MoE Question

Honest caveat before the verdict: if you simply want the best assistant a 64GB Mac can run, the dense-70B class is no longer the default answer. According to the LLMCheck index, Qwen 3.6-27B — the current Mac #1 — posts 77.2% SWE-bench Verified from an ~18 GB footprint at ~30 tok/s estimated on an M5 Max, and MoE releases like Qwen 3.6-35B-A3B (3B active parameters) are faster still. A dense 70B earns its RAM when you want its specific character: Llama's ecosystem, Hermes' steerability, or Apertus' provenance and context length. See what fits your exact Mac.

Mac Viability & Install

All three occupy roughly 40–45 GB at Q4_K_M, so a 64GB Mac is the practical floor — workable with short contexts and other memory-hungry apps closed. 96GB or 128GB gives comfortable headroom for long contexts and agent sessions, and Ultra-class Mac Studios are the ideal hosts. Smaller IQ3-class quants can squeeze under 40 GB at a quality cost.

Install: Llama 3.3 70B

ollama run llama3.3:70b
# ~43GB download, ~44GB RAM at Q4_K_M
# Context: 128K tokens

Install: Hermes 4 70B and Apertus 1.5 70B

# GGUF and MLX community builds are on Hugging Face —
# search "Hermes-4-70B" or "Apertus 1.5 70B" in LM Studio,
# or run the MLX builds directly:
pip install mlx-lm
mlx_lm.generate --model <hf-repo-id> --prompt "Hello!"

All three work with LM Studio, MLX, and any OpenAI-compatible client. Hermes 4's reasoning mode is controlled from the system prompt; Apertus 1.5's image input needs a runtime with multimodal support.

The Verdict

For most people who specifically want a dense 70B: Llama 3.3 70B. It is the known quantity — the deepest ecosystem, the most predictable behavior, one-command install. Writers, roleplayers, and anyone who wants a model that follows the system prompt rather than arguing with it should take Hermes 4 70B. And if the license question matters — commercial deployment, auditability, EU-flavored compliance — Apertus 1.5 70B is the only clean answer in the class, and it brings the longest context and image input as a bonus.

Disk is cheap and a 64GB Mac only holds one of these in memory at a time, so there is no penalty for keeping two installed and switching per task. And if none of the three profiles fits, the broader leaderboard is where the highest scores-per-gigabyte live — mostly in much smaller, much faster models.