Get AI summaries of any video or article — Sign up free
Microsoft's Phi 3.5 - The latest SLMs thumbnail

Microsoft's Phi 3.5 - The latest SLMs

Sam Witteveen·
5 min read

Based on Sam Witteveen's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.

TL;DR

Phi 3.5 mini instruct is a 3.8B, fast, locally runnable upgrade with a 128K context window and reported gains on multilingual and multimodal-related benchmarks.

Briefing

Microsoft has expanded its Phi 3 lineup with three new Phi 3.5 models—two instruction-tuned language models and an updated vision model—pushing small, locally runnable systems closer to the performance of much larger competitors. The headline is Phi 3.5 mini instruct: a 3.8B model designed for speed and local deployment that shows clear benchmark gains over the earlier Phi 3.1 mini naming, including improvements on multilingual and non-European languages, while keeping a 128K context window.

Across the reported MMLU-style results, Phi 3.5 mini instruct posts a “decent bump” over the prior mini variant, with especially noticeable gains in multimodal-related scoring and languages such as Arabic, Chinese, Thai, Japanese, and Korean. Microsoft also frames the model’s position by comparing it against larger systems—Gemma 2 9B, Gemini flash, and GPT-4o mini—suggesting that model choice should depend on the language and task mix. On long-context behavior, the model is described as capable of reaching 128K tokens for most users, with shorter contexts performing near an 8B-class baseline (Llama 3.1 8B), while longer contexts favor the larger Llama 3.1.

The second release, Phi 3.5 MOE instruct, brings a Mixture-of-Experts approach to Phi 3.5 with a much larger overall configuration: described as a 16×3.8B model with a forward path around 6.6B parameters. It’s trained on nearly 5T tokens (versus 3.4T for the mini) and also supports a 128K context window. Benchmarks are positioned as strong for an open model—MMLU results are claimed to be on par with Gemini flash and GPT-4o mini, and GSM 8K results are said to beat Llama 3.1 and Gemma 2 9B, though it still trails GPT-4o mini. The tradeoff is practicality: local use demands substantial GPU memory, with reported VRAM usage around 33+GB, implying A100-class hardware or multiple consumer GPUs.

The third model updates Phi 3 vision: a roughly 4.2B-parameter system fine-tuned on 500B tokens on top of Phi 3 mini language models. Microsoft indicates it was trained between July and August, only weeks before release. Benchmark comparisons place it ahead of some smaller proprietary options (like Gemini 1.5 flash and GPT-4o mini) while still behind Gemini Pro and GPT-4o. It’s released under an MIT license, and Microsoft pairs the release with an updated Phi 3 cookbook containing recipes for fine-tuning and end-to-end workflows using the vision model.

In hands-on testing described alongside the release, the Phi 3.5 mini instruct is the most compelling day-to-day option: it runs faster than the MOE model and performs at least as well on GSM 8K, including correcting issues seen in the MOE output. The MOE model can enable more agentic/tool-use patterns (including JSON-style outputs) but is slower and not dramatically better for the tester’s standard tasks. Overall, the new Phi 3.5 models aim to make local, private, fine-tunable assistants more capable—especially for multilingual and long-context use—without requiring the jump to proprietary, much larger systems.

Cornell Notes

Microsoft’s Phi 3.5 release adds three instruction- and vision-focused models: Phi 3.5 mini instruct (3.8B), Phi 3.5 MOE instruct (16×3.8B with ~6.6B forward path), and an updated Phi 3 vision model (~4.2B). The mini instruct is positioned as a fast, locally runnable upgrade with a 128K context window and notable gains on multilingual and multimodal-related benchmarks, including Arabic and Chinese. The MOE instruct targets higher benchmark performance but comes with heavy hardware demands (reported ~33+GB VRAM) and slower generation. The vision update is fine-tuned on 500B tokens and released under an MIT license, alongside an updated Phi 3 cookbook for vision fine-tuning and end-to-end recipes. The practical takeaway: for most local use, Phi 3.5 mini instruct is the best balance of speed and quality, while MOE is for users who can afford the compute.

What makes Phi 3.5 mini instruct stand out compared with earlier Phi 3 mini variants?

It keeps the small 3.8B footprint while improving instruction tuning and benchmark performance. Reported gains include better MMLU-style scores over the earlier mini (referred to in benchmarks as Phi 3.1 mini), with especially strong improvements on multimodal-related scoring and non-European languages such as Arabic and Chinese. It also supports a 128K context window, with shorter contexts performing close to Llama 3.1 8B and longer contexts favoring the larger model. The model is described as fast enough for local runs (e.g., via Ollama).

How do the reported language and data-training details support the multilingual claims?

Microsoft says the mini instruct was trained on 3.4 trillion tokens and supports additional languages including Thai, Japanese, and Korean. It also mentions newly created synthetic data for math, coding, and common sense. The transcript contrasts this with many open models that “aren’t supporting a lot of these languages,” implying Phi 3.5 mini’s training mix is broader than typical open releases.

What is the compute tradeoff behind Phi 3.5 MOE instruct, and why does it matter for local deployment?

Phi 3.5 MOE instruct uses a Mixture-of-Experts design: a 16×3.8B configuration with a forward path around 6.6B parameters. That architecture boosts capability but increases runtime cost. In the described tests, it consumed about 33+GB VRAM, leading to guidance that users likely need an A100-class GPU or multiple L4s (and possibly three consumer GPUs) to run it comfortably. The result is that MOE can be powerful, but it’s not the “drop-in” local option the mini is.

Where does Phi 3.5 MOE instruct land on benchmark comparisons, and what still limits it?

In Microsoft’s comparisons, it’s positioned as strong for an open model: MMLU results are claimed to be on par with Gemini flash and GPT-4o mini and even substantially beating GPT-4o mini on MMLU. On GSM 8K, it’s described as beating Llama 3.1 and Gemma 2 9B, while landing a bit behind GPT-4o mini. The transcript also notes that it remains off the pace of larger Gemini Pro and GPT-4o class models (in the broader set of comparisons).

How does the updated Phi 3 vision model differ from the language models, and what practical resources come with it?

The vision model is an updated ~4.2B-parameter system fine-tuned on 500B tokens on top of Phi 3 mini language models, with training reportedly occurring between July and August—weeks before release. It’s described as beating some smaller proprietary vision-adjacent options (Gemini 1.5 flash and GPT-4o mini) while still trailing Gemini Pro and GPT-4o. It’s released under an MIT license and comes with an updated Phi 3 cookbook containing recipes for fine-tuning and end-to-end tasks using the vision model, including examples like structured JSON output and tool-use/agentic workflows.

Review Questions

  1. Which Phi 3.5 model is most practical for local use in the transcript, and what two reasons are given for that choice?
  2. How do the transcript’s benchmark comparisons differ between Phi 3.5 mini instruct and Phi 3.5 MOE instruct?
  3. What hardware requirement is highlighted for running Phi 3.5 MOE instruct, and how does it affect who should choose it?

Key Points

  1. 1

    Phi 3.5 mini instruct is a 3.8B, fast, locally runnable upgrade with a 128K context window and reported gains on multilingual and multimodal-related benchmarks.

  2. 2

    Microsoft attributes the mini’s improvements to training on 3.4 trillion tokens plus newly created synthetic data for math, coding, and common sense.

  3. 3

    Phi 3.5 MOE instruct uses a Mixture-of-Experts design (16×3.8B; ~6.6B forward path) and targets stronger benchmark performance, but it is far more demanding to run locally.

  4. 4

    The MOE model is reported to require roughly 33+GB VRAM in testing, making A100-class or multi-GPU setups more realistic than single consumer cards.

  5. 5

    The updated Phi 3 vision model is about 4.2B parameters, fine-tuned on 500B tokens, and released under an MIT license.

  6. 6

    Microsoft pairs the releases with an updated Phi 3 cookbook, including recipes for vision fine-tuning and end-to-end workflows (including structured outputs and tool-use patterns).

Highlights

Phi 3.5 mini instruct keeps the small 3.8B size while improving multilingual performance—especially for languages like Arabic and Chinese—without sacrificing a 128K context window.
Phi 3.5 MOE instruct can deliver strong benchmark results for an open model, but the runtime cost is steep: ~33+GB VRAM is cited in testing.
The updated Phi 3 vision model (~4.2B) is fine-tuned on 500B tokens and ships with an MIT license plus a refreshed Phi 3 cookbook for vision workflows.

Topics

Mentioned

  • MMLU
  • MOE
  • VRAM
  • JSON
  • MIT