LoginSign Up
← Back to all sparks
O

Ollama

AI-ASSISTANTS
Velocity5.0

Run and manage large language models locally on your own machine.

ollama.com ↗

Ollama ships on the frontier-model release calendar, with an MLX build attached to each drop.

local inferencemodel supportmlxapple siliconquantizationcoding agents
◆Current state
Ollama's current window is almost entirely about what it can run and how fast it runs it. Qwen 3.8 27B arrives in v0.32.12 with a separately tuned MLX variant for Apple Silicon, and v0.32.13 completes that model's steering surface a day later. The rest is quantization and prefill work — NVFP4 global-scale kernel fusion for roughly 7-8% faster prefill — plus launch integrations for third-party coding harnesses. v0.32.14 is the smallest entry in the window: WebP transcoding for llama-server and a qwen renderer that no longer insists system messages come first.
◆Where it's heading
MLX is no longer a side path here. Every recent model addition lands with an Apple Silicon build tuned separately from the CUDA one, and the performance and defaults work — NVFP4 fusion, repeat_penalty matched to what other engines do — reads as Ollama closing the gap with the runtimes it gets benchmarked against rather than differentiating from them. What v0.32.14 adds to the picture is the maintenance tail: input-format and message-shape fixes arriving days behind a model launch, which is what tracking someone else's release schedule actually costs.
◆Prediction
Expect the next notable release to be another same-week model addition with a paired MLX build, since four of the last six entries take that shape, with small renderer and input-handling patches trailing it. Whether the coding-harness integrations keep accumulating is harder to call — v0.32.11 is the only entry in this window that touches them.

◆Recent moves

  1. 2d ago

    WebP images accepted; qwen tolerates late system messages

    Two narrow compatibility items: llama-server now transcodes WebP images, and the qwen renderers accept system messages that are not first in the list. Neither adds capability so much as removes a reason for a request to fail, and both are the kind of cleanup that follows a fast model-enablement run.

    View source ↗
  2. 4d ago

    Qwen 3.8 27B lands, with an MLX build for Apple Silicon

    Qwen 3.8 27B becomes runnable via `ollama run qwen3.8:27b`, with a separately optimized `27b-mlx` tag for Apple Silicon aimed at repeated coding-agent workloads. This is the arc's routine business — a frontier open model added within the release window, with the Mac path tuned as its own target rather than inherited.

    View source ↗
  3. 4d ago

    Qwen 3.8 gains developer-instruction support

    A single-item follow-up adding developer-instruction support to qwen3.8. Despite the feed timestamping it ahead of v0.32.12, its changelog range confirms it lands after the initial Qwen 3.8 27B enablement — the two are one piece of work, with the model arriving first and its steering surface completed immediately after.

    View source ↗
  4. 5d ago

    Muse Code and DeepSeek Harness launch integrations

    Two more coding harnesses — Muse Code and DeepSeek Harness — gain launch integrations, with a renderer change to match Muse Glimmer's reasoning template. Individually small, but it continues the thread of Ollama making itself the default local runtime that other agent tools launch against.

    View source ↗
  5. 6d ago

    repeat_penalty now defaults off; NVFP4 prefill ~8% faster

    The v0.32.10 final ships the NVFP4 prefill speedup plus a default change turning repeat_penalty off to match other engines, and fixes a blob verification skip when an OCI manifest's config and layer share a digest. The default change is the notable one: it aligns Ollama's output behavior with the runtimes it is measured against, at the cost of older models that relied on the 1.1 default.

    View source ↗
  6. 6d ago

    Release candidate: fused multiply-and-cast for NVFP4 prefill

    The release candidate carrying the NVFP4 kernel fusion, superseded roughly two hours later by the v0.32.10 final that delivers the same change. Useful as the engineering record — order-swapped A/B medians, byte-identical greedy outputs — but not a separate shipped release.

    View source ↗