My recent AnythingLLM/Ollama adventure on the Lenovo ThinkStation P3 Ultra focused on utility. I wanted a capable local AI setup that could answer questions without acessing the cloud. Then a simple performance question came along: “How fast is this thing, really?” Ollama makes that easy to determine. Its verbose mode reports prompt-processing and output-generation statistics after a response. Once I knew where to look, the experiment practically designed itself. On the way, the P3 Ultra shows some AI chops, too.
I ran the installed Qwen3.6 model directly through Ollama rather than through AnythingLLM. That removes the front end from the timing exercise and lets Ollama report what its inference engine is doing. I deliberately chose a subject I know well: Windows Secure Boot. The prompt asked the model to explain, in roughly 300 words, how Secure Boot validates the boot chain. That made the answer long enough to produce a useful measurement instead of a tiny burst dominated by startup effects. FWIW, it was a pretty good answer, too.
How the P3 Ultra Shows Some AI Chops
The response turned out much longer than 300 words at 1,502 output tokens, which was fine for this purpose. Ollama reported 19.924866 seconds of evaluation time and an evaluation rate of 75.38 tokens per second. Total duration was ~20.7 seconds. The prompt itself contained 28 tokens, processed at 37.72 tokens per second.
For local-LLM performance, that 75.38 tokens-per-second evaluation rate is the headline number. It represents output generation, not model loading and not the time required to consume the prompt. The arithmetic checks out, too: 1,502 tokens over 19.924866 seconds lands at essentially the same 75.38 TPS that Ollama itself reported.
More important, the run felt fast. Numbers matter because they let us compare systems, but responsiveness is what makes a local AI system usable. Seventy-five generated tokens each second is comfortably beyond the rate at which I can read the answer as it appears. For interactive question-and-answer work, the P3 Ultra was not leaving me staring at the screen waiting for the next few words.
What Exactly Was Running?
A benchmark without the model details is not much of a benchmark. The next step was therefore ollama show qwen3.6:latest. My installed model identifies itself as the qwen35moe architecture with 35.5 billion parameters and Q4_K_M quantization. It advertises a 262,144-token context length and supports completion, vision, tools, and thinking. Thinking is enabled by default in the model metadata I inspected.
The Mixture-of-Experts detail matters. Qwen’s published model information describes Qwen3.6-35B-A3B as having 35 billion total parameters with 3 billion activated during inference. In other words, the big total parameter number should not be read as though every parameter participates equally in generating every token. That architecture helps explain why a model of this overall size can deliver lively local performance.
Ollama’s library also lists the Qwen3.6 35B Q4_K_M variant in this model family. My local installation reported 22 GB in ollama list. I prefer reporting what the machine itself told me about the artifact I actually ran rather than silently substituting a current download-page size.
CPU and GPU Share the Job
The final useful clue came from ollama ps while the model was still resident. That command reported the 22 GB Qwen3.6 model running with a 19%/81% CPU/GPU split, a 4,096-token allocated context, and the llamacpp runner. Thus the 75.38 TPS result did not come from a benchmark with the model reported as 100% GPU-resident. Some of the workload was offloaded to CPU/system memory.
There is another useful distinction here. The model can support a far larger context than the 4K context allocated for this particular run. For my short Secure Boot prompt, that was no handicap. It does mean I would record 4K as part of the test conditions. Change the context allocation, model quantization, hardware placement, prompt, or software stack and the throughput number can change with them.
One Run Is a Baseline, Not a Lab
I won’t present this single pass as a rigorous benchmark suite. It is better understood as a reproducible field measure. The model was identified, the prompt was substantial, the output was long enough to time meaningfully, and Ollama supplied token counts plus generation duration. Those pieces give me a useful baseline for the P3 Ultra as I continue working with AnythingLLM and Ollama.
If I decide to turn this into formal testing, the next step is obvious: use identical prompts, repeat each run several times, control warm versus cold model state, record the exact context allocation, and compare model variants under the same conditions. For now, though, 75.38 TPS tells me what I wanted to know. This workstation can make a decent-sized local model feel fairly interactive.
A PS4 Supplies a Fun Reality Check
My question about the P3 Ultra arose after I saw a Tom’s Hardware story about Garlic LLM running Qwen3.5-9B on a jailbroken PlayStation 4. The reported performance is 31 to 39 tokens per second for text chat and up to 43 tokens per second for code. Garlic LLM accomplishes that on a base or Slim PS4 by talking directly to the console’s GPU, which is a clever achievement by itself.
I can’t call this an apples-to-apples benchmark. The PS4 uses a different Qwen model, different quantization choices, a different inference implementation, and an unpublished prompt/test method in the Tom’s account. My P3 Ultra run used Qwen3.6, produced 1,502 tokens, and gave me Ollama’s timing data directly. Still, 75.38 TPS versus a reported maximum of 43 TPS makes a perfectly good informal reality check.
More tellingly, the P3 Ultra reached that rate with Qwen3.6-35B-A3B, a model whose published specification lists 35 billion total and 3 billion active parameters. I started this exercise wondering whether my compact workstation had respectable local-AI chops. The answer looks pretty clear: Heck, yeah! The P3 Ultra shows its AI colors nicely, indeed. I’m still using it, too, even though Copilot is back. I’ll be interested to see how things play out in the weeks and months ahead.
