agents · review · first seen Oct 28, 09:33 UTC
Ollama adds Liquid AI models for Apple hardware
Ollama v0.40.12 shipped LFM3 and d1 2B support on Apple hardware. Builders can test a text model and a decision model in Ollama, but the release does not establish how either performs on a particular workflow.
In short
- Ollama v0.40.12 shipped lfm3 and d1 2b support on Apple hardware.
- LFM3's 1.7B text model lists a 256K-token context window; that is not a workflow benchmark.
- Test the documented commands now; wait for evidence on reliability, cost, and other platforms.
What shipped
Ollama v0.40.12 shipped support for lfm3 and d1 2b on Apple hardware. The release gives two entry points: `ollama run lfm3` and `ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"`. These are starting points for evaluation, not evidence that either model meets a production threshold. Support for Linux and Windows is described as coming soon, not shipped in this release. [1]
The LFM3 family includes a 1.7B text model with a 256K-token context window. The release also describes lfm3-vl for image understanding, lfm3-audio for speech recognition, translation, and text-to-speech in 18 languages, and lfm3-omni for text, images, and audio. It presents the family as built for local agents, tool calling, and structured output. Those are stated capabilities and intended uses; the excerpt does not report measured reliability for them. [1][2]
The d1 line now has a 2B size built on LFM3. The release describes d1 as making decisions about text and images, with d1-omni adding audio. Its `--ask` example frames a billing question against a short customer message. That gives a builder a concrete decision task to probe rather than assuming a general chat response will fit the workflow. The example shows invocation syntax, not classification accuracy. [1]
A separate llama.cpp b11938 pre-release excerpt says it added LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B. It also lists LFM3 recurrent-state rollback work, an LFM3 Metal fusion baseline test, and an F32 treatment for LFM3-VL depthwise convolution kernels when quantizing. That is a distinct pre-release integration path. It does not broaden Ollama's stated Apple hardware support for lfm3 and d1 2b in v0.40.12. [1][2]
Why it matters if you build
For an agent workflow on Apple hardware, the immediate change is a model choice to test through Ollama. Start with the text path for prompts and structured outputs, then use d1:2b when the task is an explicit decision like the billing question in the release. Compare the answers with cases your workflow needs to handle. The release supplies commands and intended uses, but no pass rate. A working invocation should be the beginning of evaluation, not its conclusion. [1][2]
The 256K-token context window is a model specification. It is not a promise that a long input will be cheap, fit a particular Apple machine, or produce a reliable decision. If your workflow uses long documents, test representative input lengths and inspect the answers before designing around the maximum. The supplied material gives no memory requirement, latency, throughput, energy-use, or price figure. It cannot support a cost comparison with another local or hosted model. [1][2]
The modality list matters only if the workflow needs it. Image understanding, speech tasks in 18 languages, and combined text-image-audio use are described for named LFM3 variants, while d1-omni adds audio to the decision line. Do not read those descriptions as proof that every variant has the same Ollama support, command syntax, quality, or hardware footprint. For a text decision today, the release's explicit d1:2b example is the clearest test surface. [1][2]
How to try it today
- On Apple hardware, use Ollama v0.40.12 for this test.
- Run `ollama run lfm3` and submit a representative text prompt.
- Run `ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"` to reproduce the release's decision pattern.
- Replace the example message and question with text decisions from your workflow. Check each answer against your expected outcome.
- If long context matters, test representative inputs before relying on the stated 256K-token window.
What we checked
Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.
-
shipped L0 · primary
Ollama v0.40.12 supports lfm3 and d1 2b on Apple hardware.
per GitHub releases · ollama/ollama (L0); contradicted by GitHub releases · ggml-org/llama.cpp (L0) -
shipped L0 · primary
LFM3 includes a 1.7B text model with a 256K-token context window.
per GitHub releases · ollama/ollama (L0); contradicted by GitHub releases · ggml-org/llama.cpp (L0)
Consensus across sources
0 agree · 4 add nothing either way · 2 contradict
What is not confirmed
The supplied excerpts do not establish accuracy on billing decisions, tool-call reliability, structured-output validity, usable context length on a given device, memory use, speed, or operating cost. They do not establish that Linux or Windows support has shipped; the Ollama note says it is coming soon. The llama.cpp addition is described here as a pre-release, not a stable-release claim or evidence of Ollama support for every named modality variant. [1][2]
Who should act, who should wait
Move now if you build on Apple hardware and can run a measured local evaluation of text prompts or explicit decisions. The shipped Ollama commands make that evaluation concrete. Wait before committing a production agent, a long-context design, a multimodal path, or a Linux/Windows rollout if it depends on performance, variant support, or platform availability not established here. The evidence supports trying the models, not replacing an existing system without your own checks. [1][2]