Sber Drops GigaChat 3.5 Ultra into Open Source with Slimmer Architecture and Agent Ambitions

Russian tech giant Sber has officially released its latest flagship artificial intelligence model, GigaChat 3.5 Ultra, making it freely available to the global developer community via open source. Unveiled on July 6, 2026, the new powerhouse is optimized heavily for complex coding, multi-step math problems, and autonomous AI-agent workflows. By publishing the weights on platforms like Hugging Face, the company is directly signaling its intent to stay relevant in a fiercely competitive global open-source arena currently dominated by aggressive international players.

What makes this release particularly compelling is a drastic architectural pivot. Instead of pursuing the brute-force scaling path of its 700-billion-parameter predecessor, GigaChat 3.1 Ultra, Sber scaled things down to a more nimble 432-billion-parameter Mixture of Experts (MoE) hybrid model. According to technical documentation shared on Habr, this trimming shrunk the model's memory footprint by roughly 40% and boosted long-text processing speeds up to fourfold. It's a calculated move to lower the steep hardware barrier, letting developers deploy advanced agent logic on far more accessible enterprise infrastructure.

Linear Attention and DeepSeek Rivalry

At the core of these performance gains lies Sber’s proprietary linear attention technology. Unlike traditional attention mechanisms that repeatedly scan a full text from scratch for every newly generated word, this linear framework allows the system to compress and accumulate context progressively. It essentially retains a running summary of the data, much like how a human reads a lengthy report. According to internal benchmarks reported by 3DNews, this efficient architecture helps GigaChat 3.5 Ultra trade punches with top-tier open models like DeepSeek 3.2 despite its significantly smaller operational footprint.

Sber’s development team claims the model’s sharpness comes courtesy of over 1,500 distinct training experiments and a heavily filtered dataset prioritizing high-quality, human-created text. The result is a system capable of independently researching info, writing and executing its own code, and chaining together external API calls to finish complex back-office tasks. While the model remains free for consumers through Sber's regular web assistant, the open-source release aims to fuel a broader wave of domestic and international corporate automation.

Behind the Architecture Shift: The transition from a massive 700-billion-parameter beast to a refined 432-billion-parameter Mixture of Experts framework marks a profound philosophical change in Sber’s AI laboratory. For years, the industry narrative insisted that bigger was inherently better, forcing enterprises into an unsustainable arms race of hardware procurement. By leaning into MoE, where only a fraction of the total parameters activate for any given prompt, Sber is adapting to a global reality where compute efficiency is the true metric of survival.

The Strategy of Open-Source Diplomacy

Releasing the weights of a flagship model for free is rarely an act of pure altruism, and in Sber's case, it serves as a critical strategic lever. By open-sourcing GigaChat 3.5 Ultra on platforms like Hugging Face, the company is attempting to embed its ecosystem into the workflows of international developers who are increasingly cautious about relying solely on Western or Chinese tech stacks. Tech analysts note that providing high-performing, localized models acts as a potent recruitment tool and standardizes Sber's unique linear attention protocols across the broader dev community.

This open approach also compensates for geopolitical friction that restricts access to the absolute latest hardware accelerators. When engineering teams cannot simply throw more chips at a training bottleneck, they are forced to innovate at the algorithmic level. The development of proprietary linear attention is a direct byproduct of these constraints, effectively proving that software ingenuity can bypass physical hardware limitations to deliver competitive inference speeds.

Engineering Agentic Autonomy

Inside the enterprise sector, the real battleground has shifted from simple chatbots to autonomous agents, a domain where Sber is aggressively positioning GigaChat 3.5 Ultra. Early testers within the Russian banking ecosystem report that the model's revamped math and coding capabilities allow it to act less like an assistant and more like a junior developer. It can autonomously write a Python script, execute it in a sandboxed environment, catch its own syntax errors, and refine the output before presenting it to a human supervisor.

However, maintaining this level of autonomy requires an incredibly robust handling of long contexts. Sber’s approach of compressing and accumulating context linearly—rather than rescanning entire documents repeatedly—directly addresses the cost penalty usually associated with complex multi-step reasoning. If an AI agent has to process thousands of tokens of corporate documentation just to execute a single API call, traditional transformers become prohibitively expensive at scale, making Sber's architectural pivot a necessity for real-world viability.

Looking ahead, the success of GigaChat 3.5 Ultra will not be measured by its benchmark scores alone, but by how rapidly the open-source community adopts and fine-tunes it. As developers begin layering their own proprietary data over Sber's base weights, the model will likely morph into specialized variants for legal, medical, and industrial automation. For Sber, creating the foundation for that ecosystem secures its seat at the global table of generative AI pioneers, regardless of geopolitical headwinds.

Reading Between the Lines: While Sber’s pivot to an open-source, architecture-optimized model is framed as a triumph of engineering efficiency, it also underscores a harsh geopolitical reality. In an era where Western sanctions heavily restrict access to state-of-the-art silicon like Nvidia’s Blackwell or Hopper architectures, Russian tech giants cannot simply rely on brute-force computational scaling. Trimming a model's footprint by 40% is less of a voluntary philosophical choice and more of a technical necessity when your data centers are running on a finite, strictly rationed pool of hardware accelerators.

The Benchmark Paradox

There is also a persistent contradiction in claiming parity with global leaders like DeepSeek while simultaneously operating in a vastly different data ecosystem. AI models are only as good as the data they digest, and a model heavily tuned on Russian-centric enterprise workflows and filtered cultural datasets may face a steep uphill battle when deployed in generalized, multilingual global environments. Championing high performance on synthetic benchmarks frequently glosses over how these models behave when confronted with the messy, unpredictable realities of international corporate deployments.

Furthermore, the rush to open-source flagship weights presents a classic double-edged sword for enterprise security. By handing the global developer community a highly capable, agentic model optimized for autonomous coding and API execution, Sber is also giving bad actors a powerful tool that can be run locally, entirely offline, and free from corporate safety guardrails. The line between an autonomous agent that streamlines back-office logistics and one that systematically probes enterprise codebases for vulnerabilities is dangerously thin.

The Reality of the Agentic Future

Despite the grand promises of autonomous AI agents taking over complex reasoning tasks, the practical integration of such systems remains notoriously clunky. Corporate decision-makers are notoriously risk-averse, and the prospect of turning over actual API execution and sandboxed code generation to an AI—even one with sophisticated linear attention—requires a level of trust that current LLMs have not yet fully earned. Sber's success will ultimately depend on whether enterprises see GigaChat 3.5 Ultra as a genuinely reliable autonomous worker or merely a highly sophisticated drafting tool that still demands constant human babysitting.

It turns out that the secret to staying competitive in the global AI race isn't necessarily having the biggest budget or the most chips; sometimes, it’s just about realizing that a leaner, slightly paranoid model is much easier to feed than a multi-billion-parameter behemoth with an insatiable appetite for silicon.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn