← All digests

2026-04-02

India AI Digest — Thursday, April 2, 2026

  • Google released the Gemma 4 family — E2B (~5.1B), E4B (~8B), a 26B MoE with 4B active, and a 31B dense — under Apache 2.0, with the 31B and 26B variants taking #3 and #6 on the open-weights LMArena text leaderboard, multimodal across image, audio, and video, native training across 140+ languages, and on-device targets that include Jetson Nano and Raspberry Pi.
  • Bloomberg reported Sarvam AI is closing a $300–350M round at $1.5–1.55B valuation, led by Bessemer Venture Partners with NVIDIA, Amazon, and Prosperity7 participating alongside existing backers Peak XV, Lightspeed, and Khosla — the largest pure-play raise into an Indian foundation-model company to date and confirmation that the post-summit Sarvam-30B / Sarvam-105B release built investor conviction rather than spending it.
  • Position movements: foundation_model_capability +1 (Gemma 4, open-weights frontier); capital_availability +1 (India, Sarvam round size and investor mix); compute_infrastructure 0 (touched, predicted — Gemma 4's on-device targets reset what Indic-tuned edge work can credibly aim for, but no Indian deployment has shipped against the new ceiling yet).

Google ships Gemma 4 under Apache 2.0; the open-weights frontier moves and the edge target moves with it

Google released the Gemma 4 family on April 2, 2026 in four sizes: E2B (~5.1B parameters, 2.3B effective), E4B (~8B parameters, 4.5B effective), a 26B mixture-of-experts with 4B active parameters, and a 31B dense model. All four are Apache 2.0. The 31B variant scores ~1452 on LMArena and lands at #3 on the open-weights text leaderboard; the 26B MoE scores ~1441 and lands at #6. Context windows are 128K for the E-series and 256K for the larger two. The models are natively multimodal across image, audio, and video, with native function calling. Google reports native training on more than 140 languages. Day-zero inference support shipped for transformers, llama.cpp, MLX, ONNX, and a transformers.js / WebGPU browser demo; the E2B is positioned for Jetson Nano and Raspberry Pi class hardware. Reported instruction-tuned benchmarks include 85.2% MMLU-Pro and 84.3% GPQA Diamond on the 31B, and 60.0% MMLU-Pro / 43.4% GPQA Diamond on the E2B.

What this means. Two distinct things landed under one release.

The first is the open-weights frontier moving up. A 31B dense model at #3 on Arena open is no longer a small-model story — it is the open-weights ceiling that any subsequent open release has to clear or be measured against. The 26B MoE doing it on 4B active parameters is the more interesting architectural bet: roughly the inference cost of a 4B dense model with the quality of a much larger one, given a working router. Whether the router holds under production load distributions, Indic-script prompts, and tool-use chains is what the next several weeks of independent reproduction will show.

The second is what the E-series is built for. An E2B at 2.3B effective parameters with native audio input, multimodal reasoning, and credible instruction-following on a Raspberry Pi or Jetson Nano is the same shape NVIDIA's vision-language-agent demo on Jetson Orin Nano Super was built around three weeks later — voice-first, vision-augmented, fully local, no required cloud round-trip. The recipe was already implied by Gemma 3; Gemma 4 makes the recipe practical at quality levels closer to what application builders will actually ship.

The honest caveat is that 140+ languages in the training spec does not by itself say anything about per-language quality. Google's release materials do not break down Indic-language benchmarks. The released MMLU-Pro, GPQA, and AIME numbers are English-task scores. Until independent reproductions land Indic-MT-Bench, IndicGenBench, or comparable evaluations against Gemma 4, treat the Indic claim as "likely better than Gemma 3, magnitude unverified."

India angle.

  • Sarvam-30B and Sarvam-105B sit in a redrawn competitive frame. Both went open-weights under Apache 2.0 in early March on Hugging Face and AIKosh. Sarvam-30B's positioning was Indic-quality at a deployable size; Sarvam-105B's was frontier-class on Indic. Gemma 4's 26B MoE and 31B dense now define the open-weights bar those models are read against. The substantive question — does Sarvam-105B beat Gemma 4 31B on Indic-MT-Bench, IndicGenBench, and Indian-language Arena pairs — is the next quarter's empirical work, not a marketing question. AI4Bharat's evaluation harnesses are the cleanest existing path to that comparison.
  • AI4Bharat-class edge work has a credible base model. Indic-tuned compact models built on top of an Apache-2.0 base with native audio input change what a rural-deployment voice agent can be assembled out of. The end-to-end stack — Indic STT (AI4Bharat's IndicConformer or successor), an Indic-tuned E2B, Indic TTS — is now within range of off-the-shelf consumer hardware in a way that GPT-4-class cloud inference was never going to be at Indian unit economics.
  • SI and enterprise read. TCS, Infosys, Wipro, HCLTech, and the mid-tier SI cohort have been routing GenAI offerings primarily through closed-model APIs. An Apache 2.0 frontier-near open-weights model with 256K context and tool-use natively trained changes the build-versus-buy math on regulated-sector deployments where data residency, audit, and price stability matter. The 26B MoE is the variant most likely to enter enterprise pilots first; the 31B is the offline-fine-tunable benchmark target.
  • Compute-policy cross-read. If on-device VLA workflows become the default deployment shape for Indic consumer AI, the IndiaAI Mission GPU procurement assumption — frontier-scale training and centralised inference — shifts. Procurement that weighted heavily toward inference clusters now has to weigh a longer tail of edge-deployment subsidy, developer tooling for on-device, and reference-design certification for Indian-manufactured edge boards. Whether MeitY reads the release that way is the open question.

What this is not. Not a frontier-model release in the GPT-5.4 / Gemini 3.1 / Mythos-Preview sense. The closed-frontier and the open-frontier are still separated by a meaningful capability gap on hard reasoning, agentic chains, and computer-use benchmarks. The framing "open-source has caught up" is the wrong frame; the right frame is "the open-weights ceiling moved, and the edge-deployment ceiling moved with it."

Source: Google Developers Blog, April 2, 2026. → link

Companion: Hugging Face / Google blog, April 2, 2026. → link

Confidence: high on the release facts (variants, license, Arena ranks, on-device targets); medium on the India-angle reads, which are forward-looking and depend on independent Indic-language benchmarks that have not yet landed.

Bloomberg reports Sarvam AI closing $300–350M at $1.5B valuation; the investor mix is the signal

Bloomberg reported on April 2, 2026 that Sarvam AI is close to raising $300–350 million at a valuation of $1.5–1.55 billion, with Bessemer Venture Partners expected to lead and NVIDIA, Amazon, and Prosperity7 Ventures participating. Existing investors Peak XV Partners, Lightspeed Venture Partners, and Khosla Ventures are reported to be joining the round. Indian-outlet coverage adds HCLTech as a participant, though that line is secondary and not in the Bloomberg piece. Bloomberg notes the round could close within a week of the report. Sarvam's prior disclosed funding totals roughly $54 million across earlier rounds; the company released the open-weights Sarvam-30B (mixture-of-experts) and Sarvam-105B at the India AI Impact Summit in February and put weights on Hugging Face and AIKosh in early March, both under Apache 2.0.

What this means. The premise rests on Bloomberg's reporting of a not-yet-closed round. Treat the $350M and $1.55B numbers as upper bounds in a reported range until a Sarvam announcement or a regulatory filing lands. The composition of the cap table is what is unusually concrete in the reporting and what carries most of the analytical weight.

Bessemer leading places a US growth-stage tier-1 investor at the helm of an Indian foundation-model raise. That is a different posture than the prior Sarvam rounds, which were anchored on Lightspeed and Peak XV with research-network backing through Khosla. NVIDIA participating is the strategic line — direct compute relationship at a moment when H100 and Blackwell allocations are the binding constraint on Indian foundation-model training. Amazon participating is a distribution and infrastructure tie, with Bedrock as the obvious downstream surface. Prosperity7 — Saudi Aramco's venture arm — adds sovereign capital from outside the US-China bilateral. The mix is unusual for an Indian AI deal at this stage; it reads as a deliberate diversification of strategic dependencies rather than a single-thesis bet.

The valuation should be read against what shipped before it. Krutrim's January 2024 unicorn round was a pre-product capital event — investors backed the founder, the platform, and the narrative; the model came later and the disclosure quality at release did not match the positioning. Sarvam's round comes after Sarvam-30B and Sarvam-105B are open-weights on Hugging Face with model cards, after AI Impact Summit demos, and after the Sarvam Startup Program credit allocation. Whatever the round tells investors, it tells the ecosystem that the Sarvam release pattern — research-grade disclosure first, capital second — is the pattern that scaled.

The post-money capital arithmetic still matters. A $350M raise with NVIDIA on the cap table buys a multi-quarter compute runway at frontier-relevant scale, and adds the institutional plumbing for the next training run to be larger than Sarvam-105B without a second financing event. Whether the next model is a 200B+ dense, a larger MoE with a bigger active count, an Indic-tuned multimodal that pushes on the edge surface Gemma 4 just expanded, or all three is the open product-strategy question.

India angle. This is primarily a capital-ecosystem read, with a narrower set of stack implications.

  • Capital availability for Indian deep-tech AI. A $350M round into a pure-play foundation-model lab — no consumer app revenue, no SI services backbone — is a different ecosystem signal than the Krutrim or the Neysa rounds. Krutrim was founder-and-distribution-led; Neysa is infrastructure-and-PE-led. Sarvam is research-team-led with a strategic-partner cap table. For other Indian foundation-model founders fundraising in 2026, the bar Sarvam set is now not "be a credible founder with a platform" but "ship open-weights research-grade releases that move benchmarks before the round."
  • Compute strategy. NVIDIA's direct participation, on top of Sarvam's prior IndiaAI Mission GPU allocation through Yotta, hard-codes a particular path through the compute constraint: domestic data-centre tenancy plus direct chip-vendor relationship. Whether other Indian labs can replicate that posture without the Sarvam-class shipping record is the open question. AI4Bharat sits at IIT Madras and operates on a different model; Krutrim's compute story has been Ola-internal; BharatGen runs through public-sector consortia. Sarvam's pattern is now a fourth model of how an Indian frontier lab gets compute.
  • Sovereign-capital cross-read. Prosperity7's participation routes Saudi sovereign capital into Indian sovereign-AI work. The framing "Indian sovereign AI" gets more complicated when the capital stack includes a non-Indian sovereign; how that lands politically — particularly in the IndiaAI Mission's "indigenous" framing — is a question worth watching as the round closes and the cap table is published.
  • What this does not move. Indic-language capability, foundation-model capability, and research output dimensions are touched by the release pattern Sarvam executed before the round, not by the round itself. A capital event does not move a capability dimension. The capability test remains what the next model ships and how it benchmarks.

What this is not. Not validation that Indian AI has arrived. A pre-close round at a $1.55B valuation is investor conviction in Sarvam's specific execution, not a market verdict on the Indian AI stack. The framing that treats this as a sector inflection collapses the Sarvam-specific signal — research-grade releases, strategic cap-table construction, multi-quarter compute runway — into a sector claim it does not on its own support. The sector claim, if it holds, will be visible in the next two or three Indian foundation-model rounds and what they price against.

Source: Bloomberg, April 2, 2026. → link

Confidence: medium — the round is reported as imminent but not closed; Bessemer-lead and NVIDIA / Amazon / Prosperity7 participation come from Bloomberg sourcing; the HCLTech line and the $54M prior-funding total come from secondary Indian outlets and have not been confirmed against a primary Sarvam communication or a filing.

A two-item digest after a serious search across primary feeds (Sarvam, Krutrim, AI4Bharat, MeitY, IndiaAI Mission, RBI, SEBI), tier-3 global lab posts, and Indian secondary outlets for the ±2-day window. The Anthropic Mythos / Project Glasswing announcement is dated April 7 and falls outside the window — it will be the anchor for that day's digest. Several India-side stories from the week (Q1 funding-report compilations, accelerator applications) are tracker / window pieces rather than dated events and were not included.