Guide · 2026-08-05

Ollama Update — June 2026: What's New

Last updated 2026-08-05

The MLX engine matured, Gemma 4 landed, and v0.31.1 made Apple Silicon nearly 90% faster. Here's what changed in June 2026 and whether to upgrade.

By Mohamed Meguedmi · 8 min read

What actually shipped in June 2026

June was a consolidation month for Ollama. The 0.30 line opened in mid-May and spent June hardening the three big architectural bets it introduced: the MLX engine, Vulkan acceleration, and a rewritten GGUF import path. The month closed with v0.31.1 on June 30, a point release that delivered one of the largest single-version speedups of the year for Apple Silicon users.

If you last updated in May, the practical picture is this: Ollama is no longer a thin llama.cpp wrapper. It now routes models to different backends depending on your hardware, and the June releases are where that routing became reliable enough to trust in production.

Version	Date	Headline change
v0.30.x line	Through June	MLX engine, Vulkan backend, GGUF import rewrite, tool-calling stabilization
Gemma 4 12B added	Jun 15	16GB-friendly multimodal model, 77.2 MMLU-Pro
v0.31.1	Jun 30	Gemma 4 ~90% faster on Apple Silicon via multi-token prediction

For the full canonical list, the official Ollama releases page is authoritative; this guide focuses on what the June changes mean for real hardware.

The MLX engine and the Apple Silicon jump

The most consequential June change is the maturation of the native MLX engine. Instead of running every Mac model through llama.cpp's Metal path, Ollama can now hand supported models to Apple's MLX framework, which is tuned for the unified-memory architecture of M-series chips.

The payoff landed on June 30. Per Ollama's own benchmarking, Gemma 4 generates tokens nearly 90% faster on average on Apple Silicon after v0.31.1, measured across a coding-agent workload. The mechanism is multi-token prediction (MTP): the model drafts several tokens per forward pass and verifies them, cutting the per-token overhead that dominates decode on memory-bandwidth-bound hardware.

What this means in practice: on an M3 Max with 36GB, a workload that previously decoded around 40 tokens/sec now clears the mid-70s for Gemma 4 12B at Q4. That moves a 12B multimodal model from "usable" to "genuinely snappy" for interactive coding. If you run agentic loops where latency compounds across dozens of turns, this is the single biggest reason to update this month.

Bottom line: if you're on Apple Silicon and running Gemma 4, v0.31.1 is not optional. The MTP speedup is free and requires no config change beyond the update itself.

Vulkan, GGUF, and broader GPU support

June wasn't only a Mac story. The 0.30 line's Vulkan backend stabilized through the month, giving AMD and Intel GPU owners a first-class acceleration path that no longer depends on ROCm being perfectly configured. For anyone who fought AMD driver stacks in 2025, this is the quality-of-life fix that finally makes non-NVIDIA local inference practical.

The second under-reported change is the GGUF import rewrite. Importing a raw GGUF via a Modelfile is faster and more forgiving of quantization variants, and the metadata parsing is stricter — which is good long-term but means a handful of older community GGUFs will prompt a re-import on first run after upgrading.

Tool calling also firmed up. The 0.30 line tightened the JSON schema Ollama expects and returns, aligning it more closely with the OpenAI-shaped function-calling format most frameworks target. If you built against a looser pre-0.30 tool-call contract, test your integration before rolling to production. You can browse which models expose native tool calling in our model catalog.

Gemma 4 12B: the model to grab in June

The standout model addition of June is Gemma 4 12B. It's multimodal, fits comfortably in 16GB of VRAM or unified memory at Q4_K_M, and posts a 77.2 on MMLU-Pro — competitive with models twice its size. Combined with the v0.31.1 MTP speedup, it's the best capability-per-gigabyte pick added this month for developer workstations and single-GPU machines.

Model	Quant	Approx. VRAM	MMLU-Pro	Best for
Gemma 4 12B	Q4_K_M	~9.5 GB	77.2	Multimodal + coding on 16GB hardware
Gemma 4 12B	Q8_0	~13 GB	77.9	Max quality on 16–24GB
Gemma 4 4B	Q4_K_M	~3.5 GB	68.4	Laptops, 8GB GPUs, edge

The published numbers come from Google's model documentation; see the Gemma 4 12B model card on HuggingFace and the Ollama library page for exact tags and pull commands. For side-by-side scores against other June-era local models, our benchmarks hub tracks the full leaderboard.

Should you upgrade? What breaks

For the vast majority of users, June's releases are a clear upgrade. The performance gains are free, and the breaking changes are narrow and predictable.

• Apple Silicon: Update immediately. The MLX engine and v0.31.1 MTP speedup are the biggest reason to move.
• AMD / Intel GPU: Update. The Vulkan backend removes the most common source of setup pain.
• NVIDIA: Update, but the June gains are smaller for you — the flash-attention improvements for older compute-6.x cards landed later, in v0.31.2 (July 6).

What to watch: (1) A few older community GGUFs prompt a one-time re-import due to stricter metadata parsing. (2) The tightened tool-call schema can break integrations built against pre-0.30 behavior — test your function-calling flows. (3) Pin your Ollama version in CI if you depend on exact decode throughput for cost modeling, since MTP changes tokens/sec meaningfully.

If you're budgeting a local deployment against a cloud API, plug the new June throughput numbers into our cost calculator — the MTP speedup materially shifts the break-even point for Apple Silicon fleets.

How to update Ollama

1. Check your current version: run ollama --version. Anything below v0.31.1 misses the June gains.
2. macOS / Windows: download the latest installer from ollama.com/download and run it; it upgrades in place.
3. Linux: re-run the official script — curl -fsSL https://ollama.com/install.sh | sh — which pulls the newest release.
4. Verify: run ollama --version again and confirm it reports v0.31.1 or later.
5. Pull Gemma 4: ollama pull gemma4:12b, then ollama run gemma4:12b to confirm the MLX path and MTP speedup on Apple Silicon.

Programmatic consumers can track version and model metadata through the BestLLMfor public API (CC BY 4.0), or wire the numbers directly into an agent via our open-source MCP server — both expose the same benchmark data referenced in this guide.

Verdict

June 2026 turned Ollama's 0.30 architecture from ambitious to dependable. The MLX engine and the v0.31.1 multi-token prediction speedup deliver the biggest Apple Silicon gains of the year, Vulkan finally makes AMD and Intel inference painless, and Gemma 4 12B is the model most users should pull today.

Your hardware	Upgrade?	Why
Apple Silicon (M-series)	Yes — now	MLX engine + ~90% MTP speedup on Gemma 4
AMD / Intel GPU	Yes	Stabilized Vulkan backend removes ROCm pain
NVIDIA GPU	Yes (smaller gains)	GGUF + tool-call fixes; flash-attention wins come in v0.31.2
Production integration	Test first	Verify tool-call schema and pin the version for cost modeling

The recommendation is simple: update to v0.31.1 or later, pull Gemma 4 12B, and re-baseline your throughput. If you're on a Mac, the free performance alone justifies the two-minute install.