Phi-4 vs Mistral Small 3.2 for local text
A 14B and a 24B dense model for a 16 to 24 GB machine
Facts from the reviewed catalog
| Specification | Phi-4 | Mistral Small 3.2 24B |
|---|---|---|
| Task | Text & chat | Text & chat |
| Parameters | 14B | 24B |
| License | MIT | Apache 2.0 |
| Commercial use | Allowedlicense text | Allowedlicense text |
| Quality · LMArena 2026-10-22 · text | 1,256 ± 51,252–1,261 · 24,126 votes · rank 322 · source | 1,356 ± 51,351–1,361 · 17,454 votes · rank 212 · source |
| Memory | Q4 planning estimate (~11 GB of weights before the machine's reserve) | Q4 planning estimate (~19 GB of weights before the machine's reserve) |
| Checkpoint | official card | official card |
In LMArena 2026-10-22 · text, Mistral Small 3.2 24B rates 1,356 ± 5 against 1,256 ± 5 for Phi-4, 100 points ahead with intervals that do not overlap. Blind pairwise votes by people on the hosted model, CC BY 4.0. A rating is relative; overlapping intervals mean the order is not settled. Scale: Arena rating (style control, overall) ± half the confidence interval.
Memory on three machines
Each machine's own reserve is added (graphics cards 2 GB; unified and system memory 25%, 2 to 8 GB). Planning estimates, not measured runs.
| Machine | Phi-4 | Mistral Small 3.2 24B |
|---|---|---|
| RTX 4060 Ti · 16 GB VRAM | ~13 GBFast memory holds it | ~21 GBIn memory, outside the fast part · offload to PC RAM (32 GB assumed) |
| RTX 4090 · 24 GB VRAM | ~13 GBFast memory holds it | ~21 GBFast memory holds it |
| MacBook Air 13-inch M4 · 24 GB unified | ~17 GBFast memory holds it | ~25 GBLarger than its memory |
Phi-4: Q4 planning estimate (~11 GB of weights before the machine's reserve). Mistral Small 3.2 24B: Q4 planning estimate (~19 GB of weights before the machine's reserve). Formulas
Everything that runs on each machine: RTX 4060 Ti · 16 GB VRAM · RTX 4090 · 24 GB VRAM · MacBook Air 13-inch M4 · 24 GB unified
Which to choose
Phi-4 (14 billion parameters, MIT) is the lighter choice: its Q4 weights take about 11 GB, so a 16 GB card holds it. Mistral Small 3.2 (24 billion, Apache 2.0) needs about 19 GB of Q4 weights: a 24 GB card runs it, but a 24 GB Mac does not, because macOS keeps part of the memory for the system. It also reads images. Mistral Small publishes a 131,072-token context against 16,384 for Phi-4, which matters for long documents. In the same LMArena text release Mistral Small 3.2 rates higher; the intervals are shown below.
Files and run commands
The default file of the most downloaded collected source, or the commands each card documents. Not run by this site.
Phi-4
microsoft/phi-4-gguf · phi-4-IQ4_XS.gguf · IQ4_XS · 8 GB · microsoft
ollama run hf.co/microsoft/phi-4-gguf:IQ4_XSllama-server -hf microsoft/phi-4-gguf:IQ4_XSlms get https://huggingface.co/microsoft/phi-4-gguf- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Mistral Small 3.2 24B
unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF · Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf · Q4_K_M · 14.3 GB · unsloth
ollama run hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:Q4_K_Mllama-server -hf unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:Q4_K_Mlms get https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Numbers each publisher reports
Each card's own results with its own settings. Publishers choose different tests and settings, so these values are never compared across publishers here.
Phi-4 · reported by Microsoft
- MMLU
- 84.8Popular Aggregated Benchmark · simple-evals
- GPQA
- 56.1Science · simple-evals
- HumanEval
- 82.6Code Generation · simple-evals
- SimpleQA
- 3.0Factual Knowledge · simple-evals
- DROP
- 75.5Reasoning · simple-evals
Values as printed on the card, with the publisher's own settings (card section: Model Quality); column “phi-4 (14B)”. Not comparable across publishers and not used in any ranking. Model card
Mistral Small 3.2 24B · reported by Mistral
- Wildbench v2
- 65.33%Instruction Following / Chat / Tone
- Arena Hard v2
- 43.1%Instruction Following / Chat / Tone
- IF (Internal; accuracy)
- 84.78%Instruction Following / Chat / Tone
- Infinite Generations (Internal; Lower is better)
- 1.29%Infinite Generations
- MMLU
- 80.50%STEM
- MMLU Pro (5-shot CoT)
- 69.06%STEM · 5-shot
- MATH
- 69.42%STEM
- GPQA Main (5-shot CoT)
- 44.22%STEM · 5-shot
Every reported result
Values as printed on the card, with the publisher's own settings (card section: Benchmark Results › Text › Instruction Following / Chat / Tone; Benchmark Results › Text › Infinite Generations); column “Small 3.2 24B Instruct”. Not comparable across publishers and not used in any ranking. Model card
Limits and dates
The published quality benchmark does not test a local quantized build. Memory estimates are not speed measurements or guaranteed hardware fits. Repository size, when available, is not VRAM. Formulas
Page updated · Catalog reviewed · LMArena 2026-10-22 · text