Skip to main content
MindStudio
Pricing
BlogAbout
My Workspace
Microsoft Decision 1.1Gemma 4 fine-tuneMicrosoft AI models Ignite

Microsoft Decision 1.1: A Gemma Fine-Tune Microsoft Didn't Mention On Stage

Microsoft's Decision 1.1 model beats Google on decision benchmarks, but it's quietly fine-tuned on Google's own Gemma 4 model.

Edited by Luis Chavez-Mattos, Director of Product RSS
Microsoft Decision 1.1: A Gemma Fine-Tune Microsoft Didn't Mention On Stage

What is Microsoft Decision 1.1, and why does its base model matter?

Microsoft Decision 1.1 is one of seven AI models Microsoft announced at its Ignite 2025 event, pitched on stage as a model that beats both Google and OpenAI on a decision-making benchmark. What Microsoft didn’t say during the keynote is that Decision 1.1 is a fine-tune of Google’s own open-weight Gemma 4 27B model. The detail only shows up in a developer blog post, well past the headline claims. It’s legal and not uncommon in the industry, but it complicates the “we beat Google” framing when the win is built on Google’s own architecture.

TL;DR

  • Decision 1.1 was announced by Satya Nadella on stage as beating Google and OpenAI on the decision index benchmark, without disclosing it’s a fine-tune of Gemma 4 27B.
  • The base model disclosure only appears in a Microsoft developer blog section titled “how we built Microsoft Decision 1.1,” not on the main models page or in the Ignite keynote.
  • This isn’t a first: the earlier Decision 1.0 model was reportedly built on Alibaba’s Qwen 3 59B, meaning Microsoft’s “decision” model lineage has leaned on open Chinese and Google base models rather than homegrown architectures.
  • On artificial analysis, an independent benchmark source, Decision 1.1 does score above Google’s own Gemini-based decision models, so the performance claim itself holds up even if the branding obscures the lineage.
  • Microsoft’s broader Ignite strategy centers on the harness (the agent loop, routing, tools, and context around a model) rather than on proprietary foundation models, and its new models are explicitly positioned for cheaper, shorter tasks rather than frontier reasoning work.
  • Because Gemma 4 is released under an Apache 2.0 license, Microsoft’s fine-tuning and repackaging is fully permitted; the issue is disclosure and framing, not legality.
Remy Certified AI Engineer badge
Live with Dmitry ShapiroEx-Google product lead · Ex-MySpace CTO

What you'll learn · Dec 2

  1. 01How enterprises really build with AI
  2. 02What full-stack means and why it's critical for business software
  3. 03Prompting that holds up in production
  4. 04Build and deploy your own app, live

Why would Microsoft build on Gemma instead of its own architecture?

Training a frontier-capable base model from scratch costs enormous amounts of compute and data curation. Fine-tuning an existing open-weight model is far cheaper and faster, and it lets a company focus its effort on the part that actually differentiates the product: post-training, reinforcement learning, calibration, and the inference stack around it. Microsoft’s blog describes doing meaningful work on top of Gemma 4 27B for Decision 1.1, including RL and calibration passes, plus what appears to be a custom inference engine.

This is a common pattern now. Open-weight releases from Google (Gemma), Meta (Llama), and Alibaba (Qwen) give any company with enough fine-tuning expertise a shortcut to a competitive model without owning foundation-model research. Microsoft has used this shortcut twice in its decision-model line: Decision 1.0 reportedly sat on top of Qwen 3 59B, and Decision 1.1 sits on top of Gemma 4 27B. Both are allowed under their respective open licenses. The friction is that Microsoft presented Decision 1.1 as beating Google without noting that Google’s own model is the foundation underneath it.

Does Decision 1.1 actually beat Google and OpenAI?

According to artificial analysis, an independent benchmarking source, Decision 1.1 does score above Google’s own Gemini-based decision models on the decision index. That’s a genuine result, not an invented one, and it’s one of the few claims from Microsoft’s Ignite keynote that comes from a third party rather than Microsoft’s internal Evals platform.

That matters because most of the other headline numbers from Ignite, including the claim that Microsoft’s models hit 96% of Claude Opus 5.5’s performance at 5% of the cost, came from Microsoft’s own benchmarking tools rather than independent evaluators. Those internal numbers also softened under scrutiny: the 5% cost figure relied on a 50% promotional rate, and a comparison against GPT 6.1 shifted from “40% of the cost” to roughly 80% once that promotional pricing was factored out. The decision index result stands apart because it’s externally verified, even though the underlying model’s origin wasn’t disclosed on stage.

Why does this fit Microsoft’s broader harness-first strategy?

The Decision 1.1 situation lines up with a pattern visible across the rest of Ignite. Microsoft spent roughly 12 minutes of a two-and-a-half-hour keynote actually talking about models. Everything else focused on the harness: routing, context management, agent orchestration, governance, and tooling.

The best coding result Microsoft showed all week didn’t come from a Microsoft model at all. It came from Copilot Fusion, a routing system that coordinates multiple models rather than being one itself. Fusion Max scored 71.2% on Terminal Bench 4.1, ahead of Opus 5.5 and Gemini-class competitors, but the top tier of that system runs on Anthropic’s Claude Sonnet 5.5, not on Microsoft’s own MAI Code 2 model (which Microsoft placed in Fusion’s standard, not max, tier).

4 hours from idea to a live business app.

Free live class · Wed, Dec 2
Remy Certified AI Engineer badge, Cohort 01

Satya Nadella described this directly in the keynote, saying results come from “tuning the product, the harness, the tools, and the model system all as one optimization.” Microsoft’s own models, including MAI Code 2, were reportedly post-trained inside the GitHub Copilot harness itself, meaning they’re shaped specifically to work well inside Microsoft’s ecosystem rather than as general-purpose, portable models. Against that backdrop, Decision 1.1 reads less like a from-scratch model and more like another harness component: Microsoft didn’t need to build a new foundation model to win a benchmark, it needed a well-tuned layer on top of someone else’s open weights.

Is it a problem that Microsoft didn’t disclose the Gemma base model?

Nothing about this violates Gemma’s Apache 2.0 license, which explicitly permits commercial fine-tuning, repackaging, and rebranding. Microsoft isn’t breaking rules. The concern is more about framing: announcing a model as beating Google without mentioning that the model is a derivative of Google’s own release creates an impression of independent capability that isn’t quite accurate.

It also marks a shift in posture. Microsoft has historically been a meaningful contributor to open-weight research and tooling. Quietly building commercial products on top of other labs’ open releases, without crediting them on stage, is a different kind of relationship with the open-model ecosystem: extractive rather than contributive. Developers evaluating Decision 1.1 should know they’re essentially evaluating a tuned Gemma 4 derivative with Microsoft’s post-training and inference layer on top, not a ground-up Microsoft architecture.

What should developers take away from this?

The practical lesson isn’t about Microsoft specifically, it’s about what’s durable in AI infrastructure right now. Base models are increasingly interchangeable parts, swapped in and out as new open releases arrive. The harness around a model (the routing logic, the evals, the context management, the tool integrations) is what tends to persist and compound in value over time.

That cuts both ways. If you’re building products, investing effort into a harness that you control gives you flexibility to switch underlying models, including moving to a router like Fusion, without rebuilding everything. But if you tune your system too tightly to someone else’s harness (as Microsoft’s own coding models are tuned to Copilot), you gain performance inside that ecosystem at the cost of portability outside it. The more models become commodity fine-tunes of a handful of open-weight bases, the more the actual competitive differentiation sits in the layer above the model, not inside it.

Frequently Asked Questions

What base model does Microsoft Decision 1.1 use?

Microsoft Decision 1.1 is fine-tuned on top of Google’s Gemma 4 27B model, according to Microsoft’s own developer blog. This wasn’t mentioned during the Ignite keynote or on the main models announcement page.

Yes. Gemma 4 is released under an Apache 2.0 license, which permits commercial use, fine-tuning, and rebranding. Microsoft’s use of it for Decision 1.1 doesn’t violate any licensing terms.

Did Microsoft also use other companies’ open models for prior releases?

Yes. Microsoft’s earlier Decision 1.0 model was reportedly built on Alibaba’s Qwen 3 59B model, suggesting a pattern of fine-tuning open-weight models from other labs rather than training foundation models in-house for this product line.

Does Decision 1.1 actually outperform Google’s models?

On artificial analysis, an independent benchmarking platform, Decision 1.1 does score above Google’s own Gemini-based decision models on the decision index. That result is independently verified, unlike several of Microsoft’s other Ignite benchmark claims, which came from Microsoft’s internal Evals platform.

Why is Microsoft focusing more on harness products than new foundation models?

The enterprise platform for building software with AI.
Vendor Risk Portal✓ SOC 2✓ HIPAA✓ Audit log
VendorsApprovalsContractsReportsSettings
Active vendors1,284
In review37
High risk6
VendorTierRiskApprovalNEW
Northwind SupplyTier 1LowApproved
Acme LogisticsTier 2MediumApprove
Helios DataTier 1HighPending
Granite FreightTier 3LowApproved
Pinecrest LabsTier 2LowApprove
You → RemyAdd an approval step for new vendors

Microsoft’s Ignite announcements, including Copilot Fusion and Foundry Agent Service, suggest the company sees more durable value in owning the orchestration layer (routing, context, tools, agent loops) than in competing directly on foundation model research, where it can instead fine-tune or route between existing open and closed models.

Editorial standards

Presented by MindStudio

No spam. Unsubscribe anytime.