Founding member and leading the foundation models team @sarvamai.
Joined August 2009
- Tweets2.2K
- Following334
- Followers4.7K
- Likes6.5K
Rahul retweeted
Great work from @XunzhuoLiu and the vLLM Semantic Router team on D3. All six models, from 0.6B to 27B, now have merged vLLM recipes!
D3 ranked #1 on both Jev Decision Index 0.3.2 boards: 66.3 text and 71.3 vision (Oct. 12 snapshot).
Run multimodal decisions on your own infrastructure through /v1/systemone—with control over data residency, latency budgets and governance.
vLLM serves decision models too: the full D3 family, Clef-Flash, Clef-Omni and Liquid d1, through the System One API.
D3 returns choices, yes/no answers and scores with probabilities over text, JSON, images and videos. Useful for routing, classification and scoring without generating a long response.
In the recipe benchmark, D3 27B delivers 1.8× the requests/sec of its checkpoint server on 1× MI325X at 64 concurrent clients, using the same mixed workload.
Start with a current vLLM nightly:
vllm serve vllm-sr/d3
Recipes for all six sizes:
github.com/vllm-project/reci…
API and supported models:
docs.vllm.ai/en/latest/model…
Introducing Decision 3.0: towards open multimodal foundation decision models ⚡
Open and state-of-the-art from 0.6B to 27B: #1 on both Jev Decision Index boards in our evaluation, on a new Pareto frontier.
HF collection 🤗: huggingface.co/collections/v…
#OpenSource #Jev #LLM
Rahul retweeted
🤯RSI突破！腾讯AI混元四步炼出终端Agent神器，Qwen3.8-27B性能直接从51.7%提升至86.5%，小模型硬刚一线！
腾讯混元最新论文Terminal-RSI正式开源递归自我改进（RSI）打法：交替两轮“成功轨迹自重写 + 策略精炼”，让27B模型在终端任务上实现质变。
要点速览：
🔹方法极简却狠：Base → RSR-1（验证成功轨迹Recursive Self-Rewrite做SFT）→ RPR-1（继承critic的RL精炼）→ RSR-2 → RPR-2，两轮闭环，前一轮模型直接生成下一轮训练数据。
🔹成绩炸裂：Terminal-Bench 2.1从51.70%飙到86.52%（+34.82pp）；TB 4.0从0%→19.84%；TB Hard达79%；多基准平均大幅跃升约30个百分点。
🔹长程也破防：Long-Horizon Terminal-Bench process reward 0.21→0.37，严格完成0/46→5/46；SWR100成功率直接翻倍。
🔹技术硬核：验证轨迹重写 + critic权重继承校准 + TB-GAE信用分配 + Success-Prioritized Reward，把“部分进展”真正转化为“严格验证通过”。
🔹意义拉满：延续T1/RSR系列，证明RSI在终端Agent上可稳定爬坡，27B小模型就能冲到86%+水平，逼近甚至超越部分更大模型。
查看论文：arxiv.org/abs/2610.22192
这不是概念炒作，是真正把验证过的成功经验反复炼成能力，终端Agent又要卷起来了！
#RSI #递归自我改进 #腾讯混元 #Qwen #TerminalBench #AIAgent #大模型 #TerminalRSI #AgentRL #混元AI
💥太全面了吧！RSI资源地图来了！开源神器 Awesome RSI 资料库正式整理完成！这是一份精选的 GitHub 研究合集，专为探索递归自我改进的开发者与研究者打造！
RSI 不是普通迭代，而是 AI系统不仅提升自身能力，还能改进“产生改进的机制”本身。从自训练、记忆演化、Harness优化，到自动化AI研发、具身系统、安全对齐，全都系统梳理好了，再也不用一篇篇论文自己拼图。
核心亮点
🔹先打基础：奠基论文、形式化分析与综述一网打尽
🔹术语清晰：区分自我精炼 / 持续自改进 / 递归自改进 / RSI基底
🔹分层拆解：模型层、Harness层、多智能体、自演化框架全覆盖
🔹评测路径完整：直接自改进评测 + AI研发/智能体能力基准
🔹社区驱动、CC0开源，持续更新中（已600+ Star，增长飞快）
RSI资源主要分类
1️⃣基础与入门 — 奠基论文与术语厘清
2️⃣模型层RSI — 自训练、自奖励、合成数据、自博弈
3️⃣Harness层RSI — 提示优化、记忆演化、脚手架演化、自演化智能体
4️⃣多智能体自改进 — 协同演化与辩论机制
5️⃣编码/软件工程自改进 — 自修改编码Agent
6️⃣自动化AI研发
7️⃣具身与物理自改进
8️⃣进化与u开放式RSI
9️⃣安全、对齐与理论
🔟基准与评测 + 框架工具合集
非常适合做RSI研究和从事AI的人系统学习，快速建立全局认知。
👉 直达链接：github.com/lobehub/awesome-r…
#大模型 #算法 #人工智能发展 #AINEXT #RSI #RecursiveSelfImprovement #AwesomeList #开源AI #AI智能体 #AI安全
Rahul retweeted
Reasoning from scratch, round number 6!
An introduction (and implementation) of Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO).
00:00 Introduction
01:54 What makes a reasoning model different?
04:25 Reasoning traces and model capability
08:29 Accuracy and format rewards
11:34 Aha moments and DeepSeek-R1 training
14:41 Reasoning effort and answer length
18:38 RLHF and RLVR
23:04 GRPO vs. PPO
26:40 GRPO explained with a cooking analogy
31:43 The KL term and simplified GRPO
35:04 Loading the pretrained model
36:07 Loading the MATH training data
39:26 Sampling model responses
46:30 Computing verifiable rewards
49:55 Computing advantages
51:54 Token and sequence log probabilities
55:29 Implementing sequence log probabilities
57:37 Fixing the inference-mode error
1:02:24 Computing the GRPO loss
1:04:37 Putting the GRPO step together
1:09:19 The GRPO training loop
1:12:57 Training settings, logging, and checkpoints
1:17:24 Running training and inspecting outputs
1:19:28 Loading and evaluating checkpoints
1:22:33 MATH-500 results and training stability
1:24:05 Memory requirements and next steps
Rahul retweeted
xiaomi MiMo oss RL environments + training code
github.com/XiaomiMiMo/verl
huggingface.co/datasets/Xiao…
Rahul retweeted
Everything we have built at @Bodhan_AI and @ai4bharat is now live. 🎉
Open-weight speech, vision & translation models for India's languages, available on Hugging Face. The same stack, hosted at some of the lowest prices in the market.
🤗 huggingface.co/bodhan-ai
Read more about in our blogs:
>>> bodhan.ai/research/blogs
API end-points live now on Bodhan AI.
>>> console.bodhan.ai/
Shout-out to @SarvamAI, our inference infrastructure partners for helping us serve our models reliably and affordably at scale.
Dedicated to Pushpak Bhattacharyya Sir. His love for education and Indian language technology inspired all of it. 🙏
Rahul retweeted
Six months ago at the IITM Bodhan AI Conclave, we made a promise to build the AI stack for Indian education.
From promise to practice, the journey is taking shape.
Watch this space as we reveal more of what we’ve been building, leading up to Sep 5.
@ravi_iitm @WSAI_IITM
Rahul retweeted
Sarvam Epoch | Chapter 1
Join us for the first edition of Sarvam Epoch, our flagship AI conference, in Bengaluru.
Across the two days, we will share new model updates, launch new products, present our latest research, and host hands-on technical workshops.
Builder Edition, 30 July | Enterprise Edition, 31 July
Register here: epoch.sarvam.ai/
Rahul retweeted
Transformer’s Attention mechanism has come a long way. We’d like to thank the researchers and the engineers in the open-source community for continuing to make high-performance AI accessible. Please celebrate with us by sharing this post, tagging more contributors, and sharing anecdotes to complete the open history of Attention! (1/8)🧵
Rahul retweeted
We're thrilled to announce that we have raised $234M in the first close of our $300M Series B at a $1.5B valuation.
@HCLTech and @BessemerVP have joined us in this round, alongside continued support from @khoslaventures and @peakxvpartners
For countries and companies, sovereign control on the AI stack is no longer an optionality. Sarvam will be the partner of choice for this aspiration. The capital allows us to accelerate our momentum towards this full stack of models, compute, and deployments.
A huge thank you to our customers, partners, investors, and the Sarvam team for your trust and belief in what we are building. We’re just getting started.
Read more: sarvam.ai/announcing-series-…
Rahul retweeted
HISTORY LESSON: In 1968 the US, USSR, UK, France, and China signed the Nuclear Non-Proliferation Treaty, declaring nuclear weapons too dangerous for any more countries to build. All five already had them. Everyone else had to submit to inspections while the cohort pinky-promised to disarm eventually (they didn't lol). India refused to sign, pointing out the NPT didn't decide nukes were too dangerous to exist, just too dangerous for anyone who didn't have them by 1967. Anthropic sabotaging Claude for anyone building what they deem a "frontier model" is the same hypocrisy. The danger started, conveniently, the day after they finished.
Perhaps @dwarkesh_sp was more on point when he compared GPUs to nuclear bombs.
Rahul retweeted
The HF science team just made async RL weight sync ~100x cheaper on bandwidth, and you don't need a shared cluster anymore.
The problem: every RL step, the trainer typically has to sync fresh weights to the inference engine. for a 7B in bf16 that's ~14GB. for a frontier 1T fp8 checkpoint, that's ~1TB; in bf16 it would be ~2TB. per sync.
The insight: between two RL steps, ~99% of bf16 weights are bit-identical. at RL learning rates, the optimizer is whispering and bf16 literally cannot hear most of it. the stored bf16 bits don't change.
What they shipped in TRL: only the changed elements get encoded as a sparse safetensors file, dropped into a Hugging Face Bucket, and fetched by vLLM. on Qwen3-0.6B, per-step payload goes from 1.2 GB to 20 to 35 MB. This is exactly what we built Buckets for: S3-like object storage on the Hub, Xet-backed (so even full snapshots only transfer the changed chunks).
The cherry on top: we ran a FULL disaggregated training where:
- the trainer lived on one box
- vLLM ran inside a Hugging Face Space
- the Wordle environment ran in another Space
- weights flowed through one Hub bucket
no shared cluster. no RDMA. no VPN. no NCCL across clouds. just HTTPS and a bucket.
one GPU + a Hugging Face account is now enough to do real disaggregated RL. multi-replica inference fleets across regions become a small devops exercise, not a research project.
Full write-up: huggingface.co/blog/delta-we…
Open source RL keeps eating the moat!
Rahul retweeted
For a decade, we've made models wider and deeper—but we've barely changed how layers *talk* to each other.
Since ResNet's `x + F(x)` in 2015, the depth residual has been the only highway for inter-layer communication.
It's time to upgrade the staircase. 🧵
Rahul retweeted
India enters the open-weights AI race with its largest models pre-trained from scratch: Sarvam 105B and Sarvam 30B
@SarvamAI's Sarvam 105B and Sarvam 30B score 18 and 12 on the Artificial Analysis Intelligence Index respectively. Announced at the India AI Impact Summit 2026 and open-sourced under Apache 2.0, both are Mixture-of-Experts models trained entirely in India using compute provided under the IndiaAI Mission (@OfficialINDIAai). Both support reasoning and non-reasoning modes.
These are an improvement from Sarvam's previous model, Sarvam M (8 on Intelligence Index, 23.6B parameters), which was based on Mistral Small rather than pre-trained from scratch. Sarvam 105B has 106B total parameters with ~10B active per token and a 128K context window. Sarvam 30B has 32B total parameters with ~2.4B active per token and a 65K context window. Alongside the text models, Sarvam also announced Saaras v3 (Speech to Text) and Bulbul v3 (Text to Speech) with a focus on Indic languages.
Key takeaways in reasoning mode:
➤ Sarvam 105B scores 18 on the Intelligence Index. Among ~100B-class open-weights reasoning models, it trails GLM-4.5-Air (23), INTELLECT-3 (22), Mistral Small 4 (27), and gpt-oss-120B (High, 33). All four peers also activate more parameters per token
➤ Sarvam 30B scores 12 on the Intelligence Index. Among ~30B-class open-weights reasoning models, it trails GLM-4.7-Flash (30), Nemotron Cascade 2 30B A3B (28), Qwen3 30B A3B 2507 (22), and Qwen3 32B (17). Sarvam 30B activates fewer parameters than these peers.
➤ Sarvam 105B's relative strength is in select agentic tasks. Its agentic index of 25 places it ahead of INTELLECT-3 (20) and GLM-4.5-Air (21) despite trailing both on overall intelligence. Its GDPval index of 773 also edges ahead of GLM-4.5-Air (665). Both new models are a large step up from Sarvam M (Reasoning), which scored 8 on the Intelligence Index.
➤ Compared to peers, both models score lower on TerminalBench Hard (Agentic Coding & Terminal Use) and AA-Omniscience. Sarvam 105B scored 1.5% and Sarvam 30B scored 2.3% on TerminalBench Hard, compared to GLM-4.5-Air (20.5%) and INTELLECT-3 (9.1%). The AA-Omniscience Index is -60 for Sarvam 105B and -72 for Sarvam 30B. Both models have high hallucination rates relative to their accuracy, and both attempt to answer far more questions rather than abstaining, which drives the negative scores.
Key model details:
➤ Modality: Text input and output only.
➤ Context window: 128K tokens (Sarvam 105B) and 65K tokens (Sarvam 30B).
➤ Pricing: Currently free on Sarvam's first-party API.
➤ License: Apache 2.0.
➤ Availability: Sarvam's first-party API; weights available on @huggingface and AIKosh.
Rahul retweeted
Releasing Sarvam-30B & Sarvam-105B in FP8
FP8 quantization via ModelOpt PTQ.
Weights quantized to FP8, KV cache scales calibrated so the serving engine stores KV in FP8 at runtime.
~50% memory reduction vs BF16, meaningfully faster inference, minimal quality loss.
Available on HuggingFace:
Sarvam-30B FP8: huggingface.co/sarvamai/sarv…
Sarvam-105B FP8: huggingface.co/sarvamai/sarv…
Rahul retweeted
Sarvam 30B & 105B GGUF models are now available.
30B → optimized for low-latency, production inference
105B → built for advanced reasoning & long-context tasks
Sarvam 30B → huggingface.co/sarvamai/sarv…
Sarvam 105B → huggingface.co/sarvamai/sarv…
Rahul retweeted
𝗞-𝗺𝗲𝗮𝗻𝘀 𝗶𝘀 𝘀𝗶𝗺𝗽𝗹𝗲. 𝗠𝗮𝗸𝗶𝗻𝗴 𝗶𝘁 𝗳𝗮𝘀𝘁 𝗼𝗻 𝗚𝗣𝗨𝘀 𝗶𝘀𝗻’𝘁.
That’s why we built Flash-KMeans — an IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks.
By attacking the memory bottlenecks directly, Flash-KMeans achieves 30x speedup over cuML and 200x speedup over FAISS — with the same exact algorithm, just engineered for today’s hardware. At the million-scale, Flash-KMeans can complete a k-means iteration in milliseconds.
A classic algorithm — redesigned for modern GPUs.
Paper: arxiv.org/abs/2603.09229
Code: github.com/svg-project/flash…
Rahul retweeted
📢 Open-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages.
Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support coming soon. Links, benchmark scores, examples, and more in our blog - sarvam.ai/blogs/sarvam-30b-1…
Rahul retweeted
Today I’m sharing a new research paper that explores a new idea in mixture of experts architecture called “DynaMoE”.
DynaMoE is a Mixture-of-Experts framework where:
- the number of active experts per token is dynamic.
- the number of all experts can be scheduled differently across layers.
From my findings the best model has a descending expert scheduler, where beginning layers have the most experts and the end layer have the least (1 expert).
This removes the rigid Top-K routing used in most MoE models and improves parameter efficiency and training stability.
Paper: arxiv.org/abs/2603.01697
Rahul retweeted
Today, we're launching the Sarvam Startup Program.
AI is the era for builders. We want to compound that momentum by powering the startup ecosystem with high-quality models and tools so they can create with full agency.
The program provides credits, priority support, and production-ready infrastructure so startups can focus fully on building.
