Sarvam AI’s Open Source Bet Faces Early Adoption Hurdles

When Sarvam AI announced that it was open-sourcing its Sarvam 30B and Sarvam 105B models, it framed the move as a milestone for India’s sovereign AI push. 

Both models were trained from scratch on large datasets curated in-house and built using a mixture-of-experts (MoE) architecture designed to scale reasoning capabilities without dramatically increasing inference costs. Sarvam 30B is positioned as an efficient reasoning model suitable for real-time deployments, while Sarvam 105B is targeted at complex reasoning, coding, and agentic workflows.

The company says the models were trained entirely in India using compute provided under the IndiaAI mission and are already deployed internally across its products. Sarvam 30B powers Samvaad, the company’s conversational agent platform, while the larger 105B model runs Indus, its AI assistant built for multi-step reasoning tasks.

The release also comes with strong positioning around India’s language ecosystem. Sarvam claims its models outperform several comparable systems on Indic language benchmarks, particularly across the 22 Indian languages and multiple scripts.

But within days of the release, conversations across social media and industry circles suggested that open-sourcing the model was only the first step. Building a usable developer ecosystem around them may prove harder.

Building An India-Native Open Model Stack

Sarvam’s open source push is aimed at something bigger than model distribution.

The company is attempting to build what it calls a sovereign AI stack, spanning datasets, tokenisation, model architecture, training infrastructure, and inference systems.

Both Sarvam models use a sparse MoE architecture with 128 experts, allowing them to scale parameter count while keeping active compute relatively low. Sarvam 30B uses Grouped Query Attention (GQA) to reduce memory usage, while Sarvam 105B adds Multi-Head Latent Attention (MLA) to improve long-context inference efficiency.

The models were trained on large datasets that include code, mathematics, multilingual content, and specialised knowledge corpora. A significant share of the training mix focuses on Indian languages, an area where most frontier models still show uneven performance.

Sarvam also built its own tokeniser optimised for Indic languages. According to the company, this tokeniser requires significantly fewer tokens to encode words in languages such as Odia, Santali, and Manipuri compared to other open-source models.

The idea is simple: if developers building applications for India want strong performance across multiple Indic languages, Sarvam’s models should offer a clear advantage.

Vivek Tiwari, who leads AI at Aphelion Labs, said the team recently built a voice-first rural banking assistant using Sarvam’s stack.

The system combined Saaras ASR, Sarvam-30B reasoning, a RAG layer over loan rules, and Bulbul TTS, enabling farmers to query KCC loan balances in dialect-heavy Hindi without typical ASR translation failures.

So, in these cases, Sarvam is definitely proving its edge. But will use cases like this guarantee quicker adoption overall? Only time will tell.

As Adya.ai cofounder and CEO Shayak Mazumder points out, the competitive dynamics of the LLM market remain brutally unforgiving.

“The performance is not bad, but it is definitely not at the cutting edge,” Mazumder said. “The problem with LLMs today is that if you are not in the top five in performance, you find very limited usage.”

He added that Sarvam’s biggest differentiator lies in its Indic language focus.

“If I am speaking in Odia or Bengali or Assamese using other models, it will not work very well. But if I use Sarvam for those languages it might potentially work better.”

Even so, Mazumder believes the company still trails the frontier.

He said, “Beyond that, unless you are at the cutting edge, (the reasoning, the long test and time compute, the various capabilities that the new large action models, large reasoning models possess), Sarvam does not.”

“Now we’re moving on from those into the chain of models or into new world model techniques. [Among] all of those, I think Sarvam is at least a year and a half behind. Hopefully, they will catch up soon.”

Will Sarvam Hit An Adoption Wall?

Despite the technical ambition, early developer feedback suggests that the models are currently difficult to experiment with locally.

One of the biggest friction points is the lack of GGUF format releases, which many developers rely on to run models efficiently on local machines using tools such as llama.cpp.

Community members on Reddit pointed out that even a week after the release, usable GGUF versions were not available, making it difficult for many hobbyists and independent developers to test the models.

Others flagged missing ecosystem integrations. Another comment highlighted recurring issues with new open source model releases:

• Lack of day-one support in inference frameworks such as vLLM and others
• Broken chat templates or tool-calling parsers
• Incomplete integration with common developer tooling

When these pieces are missing, developers often simply move on to models that already work with their existing stack.

Industry experts echo this concern.

One expert pointed out that Sarvam’s release currently relies primarily on the Hugging Face format, meaning developers need to download safetensor manifests and manually integrate them into workflows. That process is manageable for teams using the models through APIs, but less friendly for the open-source community that prefers lightweight deployment formats.

Another expert confirmed that vLLM support is not yet fully available, though he remarked that the company is working on rolling it out.

The broader issue is not unusual for open source AI.

Floworks cofounder Sudipta Biswas says that packaging and accessibility often determine adoption more than raw model quality.

“I think with open source, while it is great for development and adoption, you really have to have things in a pre-packaged manner, so more people can use it,” Biswas said.

He noted that many developers still default to closed-source models because they are easier to start with.

“The barrier to create that POC is much higher with the open source model, no matter how good it is.”

In other words, the challenge is not just releasing the model weights. It is lowering the friction for experimentation.

Distribution, Ecosystem, And The Question Of Competition

If Sarvam wants its models to become widely used, developers say the company needs to focus on distribution and ecosystem building.

Biswas suggests one possible path: exposing the models through third-party inference marketplaces.

Platforms such as AI model marketplaces like FLORA.ai allow developers to run models instantly through hosted endpoints rather than downloading large repositories and configuring infrastructure themselves.

“If they can open an inference endpoint… almost like a ChatGPT-kind of endpoint, that can be a game changer for distribution,” Biswas said.

The logic mirrors what happened with other successful open source models.

Models like Llama, Mistral, and Qwen gained traction not just because of their capabilities but because developers could run them easily through existing infrastructure and tooling.

For Sarvam, solving these adoption barriers could determine whether the release becomes a foundational moment for India’s AI ecosystem or simply another model announcement.

Sarvam is already trying to deepen its engagement with India’s startup ecosystem with its Sarvam Startup Program, aimed at helping early-stage builders experiment with and deploy its models in production environments. The programme offers startups credits, priority technical support, and infrastructure access.

The company has also begun hosting invite-only founder sessions to discuss building custom AI models and converting proprietary data into long-term model advantages.

However, there is also a strategic clock ticking.

Global AI companies are increasingly turning their attention to multilingual and regional language models. If a player like Google releases a strong Indic-optimised model integrated directly into its cloud ecosystem, Sarvam could face a steep competitive challenge.

For now, however, the company has achieved something important.

India now has homegrown open source models built end-to-end domestically, trained on local infrastructure and optimised for Indian languages. The next challenge is far harder: turning those models into tools that developers actually use.

The post Sarvam AI’s Open Source Bet Faces Early Adoption Hurdles appeared first on Inc42 Media.