Sarvam’s Indus AI Drops the Mic

Sarvam, a Bengaluru‑based startup, launched its Indus AI chat app on February 21 2026, instantly challenging the dominance of US‑origin LLMs. Indus AI runs on a 7 B‑parameter Mistral‑compatible backbone, supports a 128 k context window, and posted benchmark scores of ARC‑AGI‑2 88.9, SWE‑bench 78 % and HumanEval 81 %—all ahead of comparable 10‑B‑plus models. The pricing model is aggressive at $0.28 per million input tokens and $0.42 per million output tokens, undercutting OpenAI’s $1.25/$10 tier and making the model attractive for developers seeking low‑cost, high‑performance inference.

The launch also showcases Sarvam’s multilingual focus: Hindi, Telugu and Marathi are baked into the tokenizer, a strategic move that fills a regional market gap left by global giants. Combined with open‑weight availability and prompt‑caching‑like KV reuse, the platform lets teams fine‑tune locally and host the model on edge devices. This mix of cost, language coverage, and technical reuse could accelerate AI adoption across banking, education and e‑commerce in India, especially where per‑hour cloud fees remain prohibitive.

Prompt Repetition: The Right Way to Ask the Same Question

Lance Eliot’s Forbes piece finally quantifies the impact of repeating a prompt: factual correctness improves by 12 % on complex math queries and code‑generation accuracy jumps 8 % when the loop is limited to three iterations. The technique works only when a confidence‑checking token is placed early, temperature is stepped down (e.g., 0.7 → 0.5 → 0.3), and the repeated prompt is embedded as a sub‑task rather than a verbatim echo.

Mechanically, the repetition restores the model’s attention to the original intent, counteracting token‑distance decay that normally drifts the answer. Practitioners can wrap user queries in a lightweight meta‑prompt that triggers the loop only when ambiguity remains, then feed the intermediate reasoning steps back into the system. Early adopters in fintech have reported a drop from 14 % incorrect answers to under 5 % after adding this guardrail, proving the approach can be cost‑effective without inflating token volume.

The Infrastructure Shift: GGML, llama.cpp and Hugging Face

On February 20 2026 GGML and llama.cpp became official Hugging Face projects, securing funding while preserving open‑source licenses. This partnership means every Hugging Face model repository ships pre‑compiled llama.cpp binaries, allowing developers to run a 7 B model on a Raspberry Pi in seconds. The move also enables custom layer‑fusion and mixed‑precision pipelines that bypass expensive GPUs, democratizing high‑performance inference for small‑team startups.

Karpathy’s exploration of “highly bespoke software development” highlights how the community is already experimenting with these low‑level primitives to build tailored agents for niche domains. Coupled with recent open‑source releases such as MiniMax’s M2.5 frontier model, the GGML‑llama.cpp stack provides a unified, standards‑compliant path for on‑prem deployment, cutting latency, eliminating per‑token cloud costs, and reducing data‑privacy concerns for enterprises that cannot ship data abroad.

Prompt Caching 201: OpenAI’s Latency‑Saving Blueprint

OpenAI’s Prompt Caching 201 guide, released on February 20 2026, details KV‑cache reuse across identical prompt prefixes. The router stores hidden‑state vectors for each unique token sequence; a new request that starts with a previously seen prefix re‑uses those vectors instead of recomputing them. Three core strategies boost cache hit rates: (1) keep the system prompt short; (2) reuse the same user prefix across a session; and (3) enforce deterministic ordering of messages.

Production tests show a 10‑million‑token workload now costs $6.20 for input and $31 for output—a 45 % reduction—and latency drops from 200 ms prefill to 70 ms. Multi‑modal requests share the same KV cache when the textual prefix matches, further conserving compute. For developers, structuring API calls to repeat the same system preamble can deliver 30‑40 % cost savings without sacrificing throughput, and the guide recommends a fallback eviction policy to keep stale state from corrupting newer sessions.

Benchmark Battlefield: What the Numbers Say

The Onyx leaderboard released on February 17 2026 side‑by‑side evaluates every major LLM on 12 benchmarks. Claude Opus 4.6 leads MMLU (92.0) and ARC‑AGI‑2 (85.2); GPT‑5 tops SWE‑bench (78 %) and Humanity’s Last Exam (91 %); Gemini 3 Pro scores 94 on MMI‑Pro and 100 % on the visual MMMLU subcategory. DeepSeek R1 shines on multilingual tasks (MMMLU 87.5) but trails on code generation (SWE‑bench 70 %).

• Pricing: GPT‑5 $1.25/M input, $10/M output; Claude Opus 4.6 $15/M input, $75/M output; Gemini 3 Pro $0.28/M both directions; DeepSeek R1 $0.28/M input, $0.42/M output.
• Parameter counts: GPT‑5 (1 T), Claude Opus 4.6 (1 T), Gemini 3 Pro (1 T), DeepSeek R1 (671 B), Llama 3 405 B (405 B), MiniMax M2.5 (397 B).
• Task specialization: Gemini 3 Pro excels on reasoning and multilingual Q&A; Claude Sonnet 4.6 is the cheapest general‑purpose option; DeepSeek V3 offers strong visual reasoning despite a lower parameter count.

Manufacturing AI: Agentic Models on the Shop Floor

The Amiko Consulting report on February 14 2026 shows Anthropic’s Claude Opus 4.6 now controls process parameters autonomously, cutting defects by 18 % in pilot factories. MiniMax’s low‑cost M2.5 frontier model is being deployed across Indian garment lines, letting operators host a 397 B LLM on a modest edge server for predictive‑maintenance queries.

Two macro‑trends accelerate this shift. First, a $650 billion global AI infrastructure investment announced early 2026 funds data‑center expansion and edge hardware, giving manufacturers the compute needed for agentic workloads. Second, “self‑validating AI” inserts a secondary verification pass before any control output is committed, preventing error‑accumulation cascades. By combining open‑source inference (GGML + llama.cpp) with these safety nets, factories can run reasoning agents locally, reduce latency, and avoid costly hallucination‑driven downtime.

Note: The information in this article might not be accurate because it was generated with AI for technical news aggregation purposes.