MiniCPM M1 1105: Does a 350M Model's Thinking Dial Actually Work?
A 229MB reasoning model with five thinking levels was tested locally on sarcasm, texting, and battery math. Here's what the dial actually changed.

What is MiniCPM M1 1105?
MiniCPM M1 1105 is a small reasoning model fine-tuned from Liquid AI’s LFM2.5 350M base. At Q4_K_M quantization it weighs about 229 megabytes, small enough to run on a laptop CPU or a phone-class device. Its headline feature is an adjustable “thinking effort” dial with five settings: off, low, medium, high, and max. Instead of always reasoning the same amount, the model is supposed to scale how long it thinks before answering, based on which setting you pick. It’s built for lightweight jobs like turning spoken commands into tool calls, following short instructions, or running fully offline where privacy matters, not for coding or deep research.
TL;DR
- MiniCPM M1 1105 is a 229MB Q4_K_M quantized reasoning model fine-tuned from Liquid AI’s LFM2.5 350M, small enough to run locally on modest hardware.
- The model exposes a five-step thinking dial (off, low, medium, high, max) that is supposed to let users trade speed for reasoning depth.
- In hands-on local testing across three prompts, the dial produced barely any measurable difference in output quality between medium, high, and max.
- On a simple battery-math word problem, every single effort level got the answer wrong, including the ones that “thought” the longest.
- Higher effort settings often violated explicit formatting rules (sentence counts, emoji counts) that the “off” setting followed correctly.
- VRAM usage stayed around 1GB, and inference speed held steady near 68 to 69 tokens per second regardless of thinking level.
- The practical takeaway is that a thinking dial on a model this small doesn’t reliably convert extra compute time into better reasoning.
How does the thinking dial work?
The dial is a setting exposed at inference time that tells the model how much internal reasoning to generate before producing a final answer. At “off,” the model skips any visible reasoning and answers immediately. At “low” through “max,” it generates a reasoning block first (visible in testing as a “think” section), then produces the final response. The idea mirrors a pattern seen in other small reasoning models: let the user decide whether they want a fast, cheap answer or a slower, more deliberate one, without needing separate model checkpoints for each behavior.
In practice, during local testing with llama.cpp, the time cost of moving up the dial was small. Off answered a sarcasm-style prompt in about 0.61 seconds. Low, medium, high, and max all landed around 2 to 2.2 seconds, with token throughput staying close to 68 to 69 tokens per second across all four thinking levels. So the dial does change latency, just not dramatically, and the question is whether that extra second or two of “thinking” buys anything in return.
What happened when the dial was tested on real prompts?
Three prompts were used to probe the dial: a sarcastic-voice instruction-following test, a casual texting-reply test with strict formatting rules, and a basic battery-life word problem.
On the sarcasm test (“explain why the sky is blue in exactly two sentences as a sarcastic TikTok influencer”), only the “off” setting respected the two-sentence limit. Every thinking level from low to max produced more sentences and, worse, introduced factually wrong science, like claiming blue light has the longest wavelength or that the sun emits green light. The “high” setting came closest to a real explanation by mentioning Rayleigh scattering, but even that answer broke the sentence-count rule.
On the texting-reply test (write a flirty but not-desperate reply, under 20 words, exactly one emoji), “off” nailed every rule in 0.31 seconds. Low, medium, and high all spent time explicitly reasoning about the one-emoji requirement in their thinking blocks, then forgot to include an emoji in the actual reply. One response left an unfilled placeholder in brackets. Only “max” got the emoji count right, but the resulting message read like a study schedule rather than a casual text.
On the battery math problem (15% battery, 5% drain per 20 minutes, 70 minutes until boarding, will the phone last?), the correct answer is no. Every single effort level, from off to max, answered yes. Only the “low” setting showed any visible division in its reasoning. The higher effort levels didn’t attempt the math at all. They just restated the question and guessed. None of the five levels included the one-sentence explanation the prompt asked for.
Is the thinking dial worth using?
Live class · Wed, Dec 2
$40,000+/mo
in SaaS spend the Remy team cut by building its own apps with AI.
We'll show you how.
Bring an idea.
Build it live.
Leave certified.
Based on this testing, not really, at least not for the kind of everyday reasoning tasks MiniCPM M1 1105 is marketed for. The pattern across all three tests was consistent: turning the dial up did not reliably improve accuracy, and it sometimes made the model worse at following explicit instructions. The “off” setting, which skips reasoning entirely, was the most rule-compliant option in two of the three tests, and the fastest in all three.
This doesn’t mean effort dials are a bad idea in general. Larger reasoning models have shown real gains from scaling test-time compute. But at 350 million parameters, the model may simply lack the capacity to use extra “thinking” tokens productively. Generating a longer reasoning block doesn’t help if the underlying model doesn’t have the knowledge or arithmetic reliability to act on that reasoning. The dial changes how long the model deliberates, not how capable it is.
Who should actually use a model like this?
Models in the 229MB to 1GB range aren’t meant to replace a capable assistant for research, coding, or multi-step problem solving. LFM-based models in particular are explicitly not positioned for coding tasks. The realistic use case is narrow: on-device voice assistants converting speech into structured tool calls, simple instruction-following in resource-constrained environments, or offline scenarios where sending data to a cloud API isn’t an option.
Even within that narrow lane, this round of testing suggests caution. A model that gets a basic battery-math problem wrong at every single thinking setting, and that drops required formatting constraints as soon as you turn reasoning on, isn’t yet reliable for tasks where correctness matters, even simple ones. The honest framing is that MiniCPM M1 1105 is a technically interesting experiment in exposing effort control to users, running comfortably in about 1GB of VRAM, but the quality of its answers didn’t track with how hard it was told to think.
Frequently Asked Questions
What is MiniCPM M1 1105 built on?
It’s a fine-tune of Liquid AI’s LFM2.5 350M model, a 350-million-parameter base model. The fine-tune adds a configurable thinking-effort dial with off, low, medium, high, and max settings.
How much hardware does MiniCPM M1 1105 need to run?
At Q4_K_M quantization, the model file is about 229MB. In local testing on a GPU it used just over 1GB of VRAM, and it’s small enough to run on CPU-only setups as well, including via llama.cpp.
Does raising the thinking level improve answer quality?
In hands-on testing across sarcasm, texting, and math prompts, raising the thinking level did not reliably improve accuracy and sometimes caused the model to break explicit formatting rules it followed correctly at the “off” setting.
Can this model do math or coding?
No. It’s not designed for coding (LFM-based models generally aren’t), and in testing it failed a basic arithmetic word problem about battery drain at every single thinking level, including max.
What is this type of model actually good for?
Small reasoning models like this one are aimed at lightweight, low-latency tasks: voice-assistant style command parsing, simple instruction following, and on-device or privacy-sensitive applications where running a large cloud model isn’t practical.



