ARC opens waitlist for its first handheld gaming device aimed at Indian gamers

New Delhi-based startup ARC has opened a waitlist for its first handheld gaming device, designed specifically for the Indian market. The waitlist went live on July 4, 2024, offering early access to product updates, launch announcements, exclusive community initiatives, and priority purchase opportunities for those who sign up at playarc.gg.

Co-founded by Jobin Joseph and Kaustubh K. Jadhav, ARC says the device is part of an integrated gaming ecosystem that includes purpose-built hardware, a proprietary gaming operating system, software services, and a community platform. The company hasn't shared any technical specifications — like screen size, processor, or battery — nor has it revealed pricing, availability dates, or launch offers. Those details remain unknown.

What ARC is up against

India currently lacks a locally-supported premium handheld from a major brand. The closest options are imported devices with no official warranty or service. The Nintendo Switch OLED sells for ₹34,999 in India but has dated hardware. The Steam Deck, ASUS ROG Ally, and Lenovo Legion Go can cost ₹60,000 to over ₹90,000 through gray-market imports, and none offer Indian support.

Android-based handhelds like the AYN Odin 2 or Retroid Pocket series provide strong emulation and Android game performance for ₹25,000-₹45,000 via import, but again carry warranty and logistics risks. The Logitech G Cloud is another cloud-focused option but relies on a weaker processor for local tasks.

ARC believes it can fill this gap. "ARC is building an integrated gaming ecosystem: purpose-built hardware, a proprietary gaming operating system, software services, and a community-led platform," the company said via its founders. The specific features of that OS and services haven't been detailed.

Implicit bets on Android and cloud

Given the absence of information about the hardware architecture, it's reasonable to assume the device will run some form of Android or a custom OS based on Android — the most common choice for non-Windows handhelds. That would support the vast library of Android games, plus emulators for retro consoles. A custom OS could also optimize for cloud gaming services like Xbox Cloud Gaming or GeForce Now, which are gaining traction in India despite latency issues.

ARC's ecosystem pitch mirrors what Nintendo and Valve do with their own software layers, but building a proprietary OS from scratch is a multi-year engineering effort. Most existing Chinese handhelds simply use an Android launcher. ARC hasn't confirmed whether its OS is a deep fork or a themed launcher.

Founders Jobin Joseph and Kaustubh K. Jadhav have not publicly detailed their background in gaming hardware or product development, and the startup's funding status is unknown. Hardware manufacturing in India is capital-intensive, requiring supply chain relationships for PCBs, batteries, and screens that small startups find hard to secure.

Analysis

ARC is targeting a real gap — no locally-supported, premium handheld exists for Indian gamers who want dedicated controls for Android games and emulation without importing. But the company is essentially vaporware until it reveals concrete specifications, pricing, funding, and a timeline.

The biggest risk is pricing. If ARC charges under ₹15,000, it will need to cut corners on components and compete with smartphones plus clip-on controllers. If it goes above ₹30,000, it enters the territory of gray-market Steam Decks and ROG Allys, which offer PC-level performance. There is a narrow sweet spot around ₹20,000-₹25,000 where custom Android handhelds like the Odin 2 have found success overseas, but those lack local support. ARC has the chance to offer warranty and service — but only if it can manufacture and distribute at volume.

For now, the waitlist is a demand test, not a product announcement. Enthusiasts should temper expectations until ARC shows it can deliver hardware that competes on performance, build quality, and price with imported alternatives it claims to replace.

Read More

ARC Opens Waitlist for First Handheld Device, Targeting a Vertically Integrated Gaming Ecosystem for India

On July 4, 2024, Indian gaming hardware company ARC opened its waitlist for its inaugural handheld gaming device, marking a step toward what the company frames as the beginning of a broader, market-specific gaming ecosystem. The waitlist is now active on the company's website, offering registrants early access to product updates, launch announcements, community initiatives, and priority purchase opportunities before general availability. The move signals ARC's intent to move beyond a single product launch and toward building a sustained platform presence in the Indian market.

ARC's public positioning centers on a vertically integrated approach: purpose-built hardware, a proprietary gaming operating system, supporting software services, and a community-led platform. Rather than marketing a standalone device, the company's strategy implies control over the full user experience—from the OS layer through to software curation and community engagement. This approach suggests an effort to differentiate from global players by tailoring the stack to local conditions, such as internet infrastructure, regional game libraries, and payment systems. However, ARC has yet to release technical specifications for the handheld—including processor, screen resolution, or battery life—leaving hardware capability unverified.

The waitlist launch from New Delhi places ARC within a competitive global portable gaming landscape, where established players such as Valve, Nintendo, and Asus already hold significant market share. ARC's differentiation hinges on its India-specific focus and the claim that its ecosystem includes software services and community features tailored for the region. The decision to offer priority purchase via waitlist suggests a controlled initial release strategy, likely to manage demand and gather early user feedback. No pricing, exact release date, manufacturing partners, or supply chain details have been disclosed, making it difficult to evaluate production readiness or market positioning against existing alternatives.

Read More

Open-Source LLMs Surpass Proprietary Models with MoE Architectures

Open-Source LLMs Reach Enterprise Frontier with MoE Architectures and Specialized Capabilities

The open-source AI revolution has crossed a critical threshold. For the first time, leading open-source Large Language Models no longer merely emulate their proprietary counterparts; they are beating them on specialized benchmarks while slashing operational costs. This shift, driven almost entirely by the Mixture-of-Experts (MoE) architecture, is turning enterprise AI from a vendor-locked cloud service into an infrastructure asset.

The MoE Takeover: Why 6 of 7 Top Open Models Share One Architecture

The data is unequivocal. Among the seven most impactful open-source model releases in the past twelve months, six have been built on the Mixture-of-Experts design. MoE enables models with trillions of total parameters to activate only a fraction—often under 10%—for any given query. This creates a stark separation between raw capacity and compute cost, allowing enterprise engineers to deploy models with GPT-4-class reasoning on commodity hardware.

The R3 Multi-Head Latent Attention Mechanism in DeepSeek

DeepSeek's V3.2 and V4 Pro exemplify the MoE advantage. With 671 billion total parameters but only 37 billion active per token, the model matches GPT-5.1 on key coding and math tasks at an estimated 1/10th the inference cost. The secret lies in its proprietary R3 multi-head latent attention mechanism, which compresses key-value cache storage by 85% while maintaining full retrieval fidelity. This allows DeepSeek to support a million-token context window without the memory overhead that would sink a dense model.

Performance Benchmarks: Open Models Now Lead

The era of "good enough for open source" is over. New benchmarks show open-source MoE models commanding the leading edge in specialized enterprise tasks.

Kimi K2.6: 1 Trillion Parameters, 32B Active, Dominates SWE-bench Pro

Kimi K2.6, developed by Moonshot AI, is the current heavyweight champion of software engineering. Its architecture uses 256K token context windows and a dynamically routed MoE stack that achieves a 58.6% pass rate on SWE-bench Pro, surpassing GPT-5.4's 57.7%. For development teams, this means an open model now generates more production-ready code patches than the most advanced closed model available.

GLM-5 and Qwen 3.5: Specialized for Depth and Breadth

Zhipu AI's GLM-5 (744B total, 40B active) boasts a unique asymmetric context window: 200K input tokens and 128K output tokens. It scores 77.8% on SWE-bench Verified, making it the strongest open model for long-form reasoning tasks like audit report generation. Meanwhile, Alibaba's Qwen 3.5 (397B total, 17B active) prioritizes linguistic reach, supporting 201 languages with a 1M token context window—all under the permissive Apache 2.0 license. This makes it the go-to choice for multinational deployment pipelines.

The Multimodal Frontier: Llama 4's Natively Multimodal Approach

Meta's Llama 4 breaks the mold as the first natively multimodal MoE model. Unlike models that bolt on image encoders, Llama 4's architecture treats text, images, and even code as first-class tokens from pre-training. Its 10 million token context window allows it to process entire codebases, documentation libraries, and video streams in a single pass. Early adopters at enterprise-scale are using it for end-to-end code review—analyzing every file in a repository simultaneously.

The Economic and Strategic Shift: 46% Code AI-Generated

The adoption data confirms the trend: 46% of all production code is now AI-generated, and projections place that figure above 50% by late 2026. However, this statistic understates the transformation. Open-source MoE models enable enterprises to move from API-based consumption to local deployment using tools like Ollama. This delivers three strategic advantages:

• Cost reduction: Inference on dedicated hardware often costs 90% less than equivalent API calls at scale.
• Data privacy and compliance: Sensitive code never leaves the corporate network, critical for regulated industries like finance and defense.
• Full customization: Fine-tuning on proprietary codebases is unrestricted by permissive licenses like Apache 2.0 or MIT.

Forward-Looking Conclusion: The Post-API Era

The open-source LLM ecosystem has reached a Moore's Law inflection point. With MoE architectures delivering proprietary-level performance at a fraction of the cost, enterprise infrastructure teams now face a clear choice. The next 18 months will see the emergence of self-hosted agentic systems running entirely on open models, operating on sensitive data behind firewalls, and generating code that never touches third-party servers. The open frontier isn't coming; it's already here, and the smartest enterprises are deploying it today.

Read More

How Frontier-Grade Open-Source LLMs Are Rewriting the Rules of Software Engineering

The balance of power in artificial intelligence is undergoing a dramatic shift. For much of 2024 and 2025, the most advanced Large Language Models (LLMs) were largely gated behind proprietary APIs, creating a tiered system where only well-funded enterprises had access to frontier capabilities. That era is officially over. In a trend reshaping the software development landscape, a new generation of frontier-grade open-source LLMs is achieving parity with—and in many coding-specific benchmarks, surpassing—the most powerful proprietary models on the market. This isn't a minor incremental update; it represents a fundamental paradigm shift in how professional software engineering teams will build, deploy, and maintain code.

The Data Behind the Shift: Benchmarks That Speak Volumes

The claim of parity is not mere marketing hype; it is backed by hard, reproducible data. In the first half of 2026, two specific model releases have crystallized this trend into an undeniable reality for technical professionals. These models are not just "good for open-source"; they are demonstrably world-class.

The Rise of GLM-5.2 and Long-Context Mastery

Released on June 15, 2026, by Zhipu AI, the GLM-5.2 model is a 754-billion-parameter Mixture-of-Experts (MoE) architecture, activating a mere 40 billion parameters per token. This efficiency is critical, but its headline feature is a 1-million-token context window. This capability, enabled by an implementation of DeepSeek Sparse Attention (DSA), allows the model to ingest and reason over an entire massive codebase or a full technical documentation library in a single query. Zhipu AI reports that GLM-5.2 achieves coding leaderboard parity with Claude Sonnet 4, making it a direct competitor for complex, agentic engineering tasks. Its release under a permissive MIT license removes virtually all legal and commercial barriers to adoption, making it a production-ready asset for any organization.

Kimi K2.7 Code: The New Coding Champion

Just days earlier, on June 11, 2026, Moonshot AI unleashed the Kimi K2.7 Code model. This 1-trillion-parameter, 32-billion-active-parameter model was specifically optimized for coding and has achieved state-of-the-art results that surpass even the most premium proprietary competitors. Its performance on key coding benchmarks is nothing short of astonishing:

• SWE-bench Pro: 58.6% (beating GPT-5.4's 57.7% and Claude Opus 4.6's 53.4%)
• SWE-bench Verified: 80.2% (a significant jump from its predecessor, K2.5's 76.8%)
• LiveCodeBench v6: 89.6%
• AIME 2026: 96.4%
• GPQA-Diamond: 90.5%

These scores are not just incremental gains; they represent a definitive statement that open-source models can now lead on the most rigorous software engineering evaluations. Furthermore, Kimi K2.7 Code introduces a unique preserve_thinking mode, which maintains full reasoning traces across multiple turns in an agentic workflow. This is a critical innovation for building reliable autonomous systems, as it prevents the common problem of a model "forgetting" its strategic plan in the middle of a complex coding task.

Why This Is a Paradigm Shift for Engineering Teams

The implications of these advances extend far beyond simple benchmark wins. For the professional software engineer, CTO, and engineering manager, this trend fundamentally alters the strategic calculus around AI adoption.

Democratization of Cutting-Edge AI Agents

Perhaps the most profound impact is the democratization of advanced agentic capabilities. Previously, building an AI agent that could autonomously patch a Docker container or reason over a 100,000-line codebase required access to expensive proprietary APIs. Now, with models like GLM-5.2 and Kimi K2.7 Code, any team can self-host a model that is production-ready for these exact tasks. This reduces the barrier to entry, allowing smaller startups and individual developers to compete with larger, well-funded organizations.

Unprecedented Control, Privacy, and Cost-Efficiency

The permissive licenses, such as the MIT license for GLM-5.2 and the modified MIT license for Kimi K2.7, unlock a level of control that is impossible with proprietary APIs. Organizations can fine-tune these models on their proprietary codebases, ensuring perfect alignment with internal coding standards and terminology. They can run the models on their own hardware, eliminating concerns about data privacy and network latency. Critically, the cost per token of self-hosting a sparse MoE model like these can be 4 to 10 times lower than using a premium API, making high-volume internal usage—like automated code review or test generation—economically feasible at scale.

Hedging Against Geopolitical and Operational Risk

In an increasingly volatile geopolitical landscape, reliance on a single proprietary model provider creates significant operational risk. Governments can impose access restrictions, API pricing can change without notice, or a provider's strategic priorities can shift away from your specific use case. The availability of open-source alternatives that are globally usable under licenses like MIT provides a crucial hedge against single-vendor dependence. This ensures business continuity and long-term strategic flexibility, a factor that is becoming increasingly important for enterprise architecture planning.

The Road Ahead: An Accelerated Pace of Innovation

The first five months of 2026 alone saw the release of six new frontier-class open-weight models. The performance gap between open-source and premium frontier models for routine tasks has already narrowed to single-digit percentage points, while the cost advantage remains significant. Models like DeepSeek V4-Pro, already at 80.6% on SWE-bench Verified, demonstrate that the trend is accelerating. For the software engineering professional, the message is clear: the era of viewing open-source LLMs as a second-tier alternative is over. The frontier is now open, and the tools to build the next generation of autonomous, intelligent software are freely available for anyone to wield.

Read More

RadixArk's Miles: Open-Source RL Post-Training for LLMs

RadixArk's "Miles": An Open-Source Lifeline for Cost-Prohibitive LLM Post-Training

On July 1, 2026, RadixArk released "Miles," an open-source framework designed to dismantle the steep computational barriers that have historically locked enterprise-level reinforcement learning (RL) post-training behind exorbitant budgets. By unifying SGLang, NVIDIA Megatron-LM, and Ray, Miles promises a pluggable, fault-tolerant stack that slashes the infrastructure tax of scaling frontier LLMs. For an industry where 45.5% of AI decision-makers cite high costs as their primary barrier to adoption—and where the AI platforms market is projected to surge from $109.9B in 2025 to $181.3B in 2026—this release is arguably the most directly actionable tool for production teams in a year’s memory.

What is it? Miles is a unified, small-footprint framework for the RL phase of LLM training—typically the most resource-intensive stage—after a model has been pre-trained. It abstracts away the combinatorial nightmare of orchestrating rollout servers, distributed training clusters, and high-speed networking. Why does it matter? Because the process of aligning a model through RL (e.g., RLHF) currently requires a massive DevOps overhead that few organizations can justify. Miles collapses that overhead into a single PyTorch-native interface.

The Architecture: A Trinity of Battle-Tested Foundations

Miles does not reinvent the wheel; it bundles three proven open-source components with a thin, intelligent orchestration layer. RadixArk engineers focused on integration rather than innovation, solving the actual pain point of teams that struggle to stitch these tools together themselves. The stack is composed of:

• SGLang for high-throughput model rollout and inference during RL trials.
• NVIDIA Megatron-LM for distributed training at scale, leveraging tensor and pipeline parallelism.
• Ray for distributed workload scheduling, fault tolerance, and cluster management.

The Pluggable "Trainer" Interface

At the heart of Miles is a small, PyTorch-native trainer class that serves as the single entry point for the entire RL pipeline. Developers only need to implement a handful of hooks—rollout loop, reward computation, and policy update—while the framework handles data sharding, gradient accumulation, and checkpointing. This eliminates the months-long engineering time typically needed to build a stable RL training loop from scratch.

Key Technical Optimizations That Cut Costs

RadixArk's engineers implemented three concrete optimizations directly responsible for reducing computational expenditure by an order of magnitude during early tests:

1. Unified low-precision recipes: Miles automatically manages precision across the entire pipeline—rollout, training, and synchronization—using FP8 wherever possible, with fallback to BF16 for critical gradients. This reduces memory footprint by up to 50% without sacrificing model quality.
2. Mixture-of-Experts (MoE)-aware alignment: The framework intelligently routes tokens to the correct expert nodes during rollout and training, preventing the "expert imbalance" that cripples naive implementations. It synchronizes expert weights via fast NVIDIA NCCL/RDMA with zero-copy memory transfers, reducing inter-node latency by roughly 40% compared to standard NCCL collectives.
3. Built-in fault tolerance and observability: Ray's native error recovery doubles as an operational cost-saver. When a node fails mid-training—common in large clusters—Miles automatically redistributes the workload to a spare node without discarding progress, reclaiming the wasted compute that would otherwise be lost to manual restarts.

Practical Impact: Lowering the Barrier to LLM Fine-Tuning

The most significant impact of Miles is its role in bridging the gap between research and production. Previously, RL post-training for models like LLaMA-3 or Qwen required a dedicated team of distributed-systems engineers and a GPU cluster valued at several million dollars. Miles reduces technical friction to a set of configuration files, enabling teams with smaller compute budgets to experiment with agentic workflows and fine-grained behavioral control.

Accelerating Production-Ready Agentic Workflows

Because Miles handles the heavy lifting of rollout scaling and reward aggregation, it allows researchers to focus on reward shaping—the secret sauce behind capable agents. Whether teaching an LLM to use external APIs, compose multi-step tool calls, or execute code reliably, the ability to iterate quickly on RL policies becomes a competitive advantage. The framework's compatibility with SGLang ensures low-latency inference, a requirement for online learning scenarios where the model interacts with live systems.

With the AI market projected to compound at 28.7% CAGR through 2030, the pressure to deliver production-grade LLM capabilities is immense. Miles directly attacks the top barrier—cost—by providing a free, open-source solution that renders obsolete the need for proprietary RL stacks. For any team serious about deploying custom-aligned LLMs, it is no longer a question of whether to use RL post-training, but how quickly they can adopt Miles to do so.

Read More