/
Navigation
Chronicles
Browse all articles
Explore
Semantic exploration
Research
Entity momentum
Nexus
Correlations & relationships
Story Arc
Topic evolution
Drift Map
Semantic trajectory animation
Posts
Analysis & commentary
Pulse API
Tech news intelligence API
Browse
Entities
Companies, people, products, technologies
Domains
Browse by publication source
Handles
Browse by social media handle
Detection
Concept Search
Semantic similarity search
High Impact Stories
Top coverage by position
Sentiment Analysis
Positive/negative coverage
Anomaly Detection
Unusual coverage patterns
Analysis
Rivalry Report
Compare two entities head-to-head
Semantic Pivots
Narrative discontinuities
Crisis Response
Event recovery patterns
Connected
Search: /
Command: ⌘K
Embeddings: large
TEXXR

Chronicles

The story behind the story

days · browse · Enter similar · o open

Microsoft releases three Phi-3.5 models designed for basic/fast reasoning and more, available for developers to download, use, and fine-tune on Hugging Face

Microsoft isn't resting its AI success on the laurels of its partnership with OpenAI.  —  No, far from it.

VentureBeat Carl Franzen

Context & Ripple Effects

Phi-3.5 extends the small-model line that began with Phi-3 Mini’s debut, when Microsoft outlined smaller and medium-sized variants for AI application developers. Publishing downloadable, fine-tunable models makes that family directly usable beyond Microsoft-hosted services.

The subsequent Phi releases show that this was a continuing product direction: Microsoft later made Phi-4 weights available under an MIT License and expanded the family with additional reasoning models on Hugging Face.

First-order effects

  • Developers can download, run and fine-tune three Phi-3.5 variants, giving them model choices tailored to basic, fast reasoning and related workloads.
  • Microsoft broadens the practical reach of its Phi portfolio by distributing models through Hugging Face rather than limiting access to a hosted interface.

Second-order effects

  • Application teams can compare smaller Phi variants against other available models and select or customize one by latency, task fit and deployment constraints rather than relying on a single general-purpose model.
  • Rival model vendors face added pressure to pair capability claims with accessible weights or developer-friendly distribution if they want to compete for fine-tuning and local-deployment use cases.

Third-order effects

  • If this release pattern continues, AI application stacks are likely to become more hybrid: organizations will mix adaptable smaller models for bounded tasks with larger systems where broader capability is needed.
  • The Phi roadmap points to distribution and iteration cadence becoming durable competitive levers alongside raw model scale, though actual adoption will depend on developer evaluation and operating requirements.

The trend: This is one data point in the shift from a single frontier-model strategy toward portfolios of smaller, task-oriented models distributed for developer customization.

Discussion

  • @rohanpaul_ai Rohan Paul on x
    Congrats to @Microsoft for achieving such an incredible result with the just released phi 3.5: mini+MoE+vision 🤯 Phi-3.5-MoE beats Llama 3.1 8B across the benchmarks Of course, Phi-3.5-MoE a 42B parameter MoE with 6.6B activated during generation And Phi-3.5 MoE outperforms [imag…
  • @yampeleg Yam Peleg on x
    How the hell Phi-3.5 is even possible? Phi-3.5-3.8B (Mini) somehow beats LLaMA-3.1-8B.. (trained only on 3.4T tokens) Phi-3.5-16x3.8B (MoE) somehow beats Gemini-Flash (trained only on 4.9T tokens) Phi-3.5-V-4.2B (Vision) somehow beats GPT-4o (trained on 500B tokens) how? lol [ima…
  • @terryyuezhuo Terry Yue Zhuo on x
    The new Phi-3.5-Mini-Instruct falls short of the Phi-3.1-Mini-128K-Instruct (the updated Phi-3 released in June) on the BigCodeBench-Hard: Complete: 16.2 (vs. 26.4) Instruct: 14.2 (vs. 17.6) It appears that the primary focus of Phi-3.5-Mini-Instruct is on multilingual [image]
  • @reach_vb @reach_vb on x
    Let's gooo.. Microsoft just release Phi 3.5 mini, MoE and vision with 128K context, multilingual & MIT license! MoE beats Gemini flash, Vision competitive with GPT4o🔥 > Mini with 3.8B parameters, beats Llama3.1 8B and Mistral 7B and competitive with Mistral NeMo 12B > [image]
  • @altryne Alex Volkov on x
    Microsoft releases new PHI models including an MoE version and a updated vision version! > Phi-3.5-MoE has 16x3.8B parameters with 6.6B active > MoE trained on 4.9T tokens >Phi-3.5-Vision is 4.2B parameters! > enables multi-frame image understanding and reasoning > outperforms