whatmodel.sh / Phi-3-mini-128k-instruct (Apr 2024)
Phi-3-mini-128k-instruct (Apr 2024)
↗ share
+ contribute
Phi-3-mini-128k-instruct (Apr 2024) is an open-weights language model from Microsoft, released April 2024. Its published size is 3.82B parameters. Licence: mit.
Catalogue identity notes
Comparison-only. Arena entry 'phi-3-mini-128k-instruct' (text). LMArena's leaderboard table (arena-leaderboard space @6f2eeeb) labels 'phi-3-mini-128k-instruct' as 'Phi-3-Mini-128k-Instruct' (Microsoft, MIT), linking https://azure.microsoft.com/en-us/blog/introducing-phi-3-redefining-whats-possible-with-slms/. Maker repo microsoft/Phi-3-mini-128k-instruct @f3c06ae: licence tag mit. 3,821,079,552 parameters from the Hugging Face safetensors metadata (dense, so active = total). Released 2024-04: month the maker's Hugging Face repo was created (createdAt 2024-04-22). LMArena's table links the April 2024 Phi-3 announcement. The HF repo was later updated in place (June 2024), so the current download is not asserted to be this checkpoint.
$ how it compares
The LLMs ranked either side of Phi-3-mini-128k-instruct (Apr 2024), out of 828 with a score, then more from Microsoft.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 631 |  OpenHermes 2.5 Mistral 7B | 39±1 | local | ||
| 632 |  Granite 3.3 2B Instruct | 39±1 | local | ||
| 633 | Phi-3-mini-128k-instruct (Apr 2024) | 39±2 | local | ||
| 634 |  Mistral 7B Instruct v0.3 | 39±1 | local | ||
| 635 |  Vicuna 33B v1.3 | 39±1 | local | ||
| more from Microsoft | |||||
| 252 | Phi-4-reasoning-plus 14B | 59±1 | local | ||
| 261 | Phi-4-reasoning 14B | 58±1 | api $0.22/M +local | ||
| 365 | Phi-4 14B | 53±1 | api $0.087/M +local | ||
$ what it's good at
coding▁▁▁▃▆██▇▆▄▃▁ #321 of 458 · 11% of its score · 4 tests
Beats GLM-4.7-Flash, Qwen3-VL 8B Instruct, Qwen3-VL 4B Instruct here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CodeReviewQA: solution identification (hard), invariant accuracy | 45.9% ✎ | ▃··▁▂▂▃▂█▃▄▂ beats 22/47 | 4.5% |
| CodeReviewQA: automated code refinement, exact match rate | 7.7% ✎ | ▁▁▁▁▃▁▂▂▃█▅▄ beats 6/47 | 3.5% |
| CodeReviewQA: change localisation (hard), invariant accuracy | 30.7% ✎ | █▂▁▁▂▂▂▃▅▅▅▇ beats 26/47 | 3.3% |
| CodeReviewQA: change type recognition, invariant accuracy | 76.9% ✎ | ▁▁▁······▁▃█ beats 39/47 | 0.1% |
chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #363 of 393 · 77% of its score · 3 tests
Beats Gemma 4 E2B, Code Llama 70B Instruct here, though they rank higher overall. Trails Gemma 3 270M IT, SmolLM2 360M Instruct, Granite 4.0 350M here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Arena-Hard (v0.1, GPT-4-1106 judge) | 15.4% ✎ | ▂▁▁▅█▅▄▃▃▅▃▁ beats 9/60 | 32% |
| Polyrating rating (bias-aware re-fit of Chatbot Arena votes, mid-2024) | 1072.0 ✎ | ▂▁▂▁▃▇█▆▇▄▂▂ beats 34/95 | 23% |
| LMArena (text) | 1129.8 ✎ | ▁▁▁▃▃▃▃▆▆▇█▃ beats 27/379 | 23% |
other tests3 not in a task · 10% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open PT-LLM Leaderboard average (9 Portuguese tasks) | 65.7% ✎ | ▃▁▁▁▁▁▁▁▂█▅▁ beats 119/276 | 4.8% |
| EuroEval Life in the UK (English knowledge MC), test split, 5-shot, accuracy | 80.2% ✎ | ▁▁▁▁▁▁▁▂▄█▄▃ beats 65/146 | 3.5% |
| Vectara hallucination leaderboard, 2025 dataset (HHEM-2.3, factual consistency rate) | 96.9% ✎ | ▁▁▂▃▄▄█▆▃▂▂▁ beats 101/140 | 2.1% |
$ where to run it
Other ways to run it
- Run it as published · needs ~7.64 GB · from huggingface.co/microsoft (official)
uv run --with torch --with 'transformers<5' --with accelerate python -c "from transformers import pipeline; print(pipeline('text-generation', model='microsoft/Phi-3-mini-128k-instruct', revision='bb5bf1e4001277a606e11debca0ef80323e5f824', torch_dtype='auto', device_map='auto')('Once upon a time', max_new_tokens=40)[0]['generated_text'])"
Official weights and documentation: huggingface.co/microsoft ↗
$ how the score was worked out · Arena-Hard-Auto v0.1 leaderbo… 32%, Polyrating: A Cost-Effective… 23%, LMArena leaderboard dataset 23% and 4 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json