We track the agents, rank the models, benchmark the hardware, and report the news. We advocate for a future where open, accessible AI benefits everyone. Curiosity over fear. Freedom as a foundation.
Data from LMSYS Chatbot Arena, SWE-bench Verified, and agent benchmarks.
Measures reasoning, multi-modality, web browsing, and tool use on assistant tasks. 466 questions across 3 difficulty levels. HAL scaffold adds ~30 pts over bare model.
812 tasks across 5 realistic websites. Tests browser-based agents. Human baseline: 78%. Best hybrid computer-use agents now lead.
Broadest benchmark: OS shell, SQL, knowledge graphs, web shopping, and more. Catches weaknesses single-domain benchmarks miss.
Sources: arena.ai · swebench.com · Rapid Claw
Agents, models, hardware, inference speed, and open source: the developments that matter for what's actually shipping and in use.
Nous Research shipped "The Surface Release": a full native desktop app (macOS/Linux/Windows) for Hermes Agent, a web dashboard admin panel, fuzzy model picker, /undo command, NVIDIA/skills trusted tap, and Quick Setup via Nous Portal. 874 commits, 542 merged PRs, 170 contributors. 187k GitHub stars.
Read more →NVIDIA released Nemotron 3 Ultra: 550B parameters with 55B active, built for long-running agent orchestration. Hybrid Mamba-Transformer, NVFP4 quantization (5x throughput), LatentMoE, multi-token prediction. 30% cost savings on agentic tasks. Fully open weights, data, and recipes.
Read more →H Company released Holo3.1 with mobile automation (AndroidWorld 79.3%), cross-harness support, and quantized FP8/NVFP4/GGUF checkpoints. On DGX Spark, NVFP4 delivers 2x speedup, cutting step time from 6.8s to 3.3s. Sizes from 0.8B to 35B-A3B.
Read more →JetBrains released Mellum2: a 12B-parameter MoE (2.5B active) for routing, RAG, and sub-agents. Apache 2.0. 2x faster inference than similar-sized models. Designed for high-frequency tasks inside larger AI systems. It's the "focal" model in your agent stack.
Read more →Pi Agent added Claude Opus 4.8 support, --exclude-tools flag for selective tool disablement, headless Codex subscription login, and streaming-aware extension input. 60.9k GitHub stars. Pi is not another agent SDK. That's the whole point.
Read more →FuriosaAI partnered with Broadcom to develop a 3rd-gen AI accelerator using multi-die chiplet design. RNGD chip is in mass production at TSMC. TCP architecture targets agentic workloads with HBM4/4E and 2nm process. CUDA-alternative SDK ships with PyTorch compiler.
Read more →Zoho-backed Netrasemi successfully tested its A2000 AI SoC, ready for edge devices. TSMC 12nm process, targeting smart cameras and automotive. Early trials with 3 customers. Part of India's DLI scheme. A4000 server chip expected Q2 2027.
Read more →NVIDIA announced Vera Rubin NVL72 is now in full production. The platform powers "agentic AI factories" worldwide with 72 Rubin GPUs per rack, NVLink-C2C fabric, and NVIDIA's own Vera ARM CPU. Multi-anchor system design includes Intel Xeon 6 and Groq LP30.
Read more →Cerebras CS-3 delivers 2,522 tokens/s on Llama 4 Maverick (400B), more than 2x NVIDIA Blackwell. Wafer-scale engine holds entire models in SRAM, eliminating memory bottlenecks. Groq LPU follows at 549 TPS. For batch processing and large-scale inference, custom silicon is winning.
Read more →Inception Labs launched Mercury 2, the first diffusion-based reasoning LLM. Generates tokens in parallel instead of sequentially. 629 TPS (up to 1,100), 5x faster than leading fast LLMs. Matches Claude 4.5 Haiku on AIME 2025 (91.1%) and GPQA (73.6%).
Read more →Not all benchmarks measure the same thing. Here's a plain-language guide to the tests that matter.
What it measures: Blind human preference. Real users compare two anonymous responses and pick a winner. ~6M+ votes aggregated via Bradley-Terry Elo.
Why it matters: Best proxy for "which model do people actually prefer chatting with?" A 100-Elo gap ≈ 64% win rate.
What it measures: Can the agent fix real GitHub issues? 500 hand-verified Python bugs. Agent submits a patch; hidden tests verify correctness.
Why it matters: Closest public proxy for "can this agent do real software engineering work?"
What it measures: General AI assistant tasks: reasoning, multi-modality, web browsing, tool use. 466 questions at 3 difficulty levels.
Why it matters: Tests multi-step reasoning with tools. Scaffold gap is huge: bare ~45%, scaffolded ~75%.
What it measures: Web navigation agents. 812 tasks across realistic Reddit, GitLab, Shopify replicas.
Why it matters: Canonical benchmark for browser-using agents. Human baseline: 78%.
What it measures: 8 environments: OS shell, SQL, knowledge graphs, web shopping, card games, household sim, lateral puzzles.
Why it matters: Broadest benchmark. Catches weaknesses others miss. Always read per-env breakdowns.
April 2026: UC Berkeley RDI showed how eval harness quirks can inflate agent benchmark scores — great news for anyone who wants honest measurement.
Pro tip: Favor independent third-party scores (Epoch AI, BenchLM) and multi-run results. The best models shine under careful scrutiny.
The model matters, but the scaffolding matters more. Same model can swing 30-50 points depending on the framework wrapping it.
From terminal-native coding agents to self-hosted personal assistants. Here's what's actually running, what's open source, and what's worth your time.
The most-starred open agent framework. v0.16.0 shipped a native desktop app (Electron, macOS/Linux/Windows), full web dashboard admin panel, fuzzy model picker, /undo command, NVIDIA/skills trusted tap, and Quick Setup via Nous Portal. Connect to remote gateways over OAuth. 874 commits in one release.
github.com/NousResearch/hermes-agent →Not another agent SDK. It's a full coding agent harness. v0.77.0 adds Claude Opus 4.8 support, --exclude-tools for selective tool disablement, headless Codex subscription login, and streaming-aware extension input. Telegram bridge, Grok subagents, taskflow orchestration, and image generation built in.
github.com/earendil-works/pi →The self-hosted assistant dream is real — pick the secure path. NVIDIA's NemoClaw wraps OpenClaw in a hardened runtime (OpenShell), and Hermes Agent delivers the same vision with strong security defaults out of the box. Encrypted credentials and locked-down websockets are now table stakes, and the ecosystem is rising to meet them.
XDA comparison →Top computer-use model. Controls desktop, browser, and mobile (AndroidWorld 79.3%). Ships quantized FP8, NVFP4, and Q4 GGUF checkpoints for local inference. On DGX Spark, NVFP4 delivers 2x speedup, 3.3s per step. Sizes from 0.8B to 35B-A3B.
huggingface.co/Hcompany/holo31 →Built for agent orchestration. Hybrid Mamba-Transformer for long context, NVFP4 quantization (5x throughput vs BF16), LatentMoE routing, multi-token prediction. 30% cost savings on agentic tasks. PinchBench 91%, Ruler @1M 95%. Fully open: weights, data, recipes, NeMo RL training code.
huggingface.co/nvidia/Nemotron-3-Ultra →The "focal" model for agent stacks. Built for routing, RAG pipelines, sub-agents, and high-frequency text+code tasks. 2x faster inference than similar-sized models. Designed to be the fast, cheap layer between your expensive reasoning model and your tools.
huggingface.co/JetBrains/mellum-2 →NVIDIA keeps raising the bar while custom ASICs from Cerebras, Groq, SambaNova, and FuriosaAI redefine inference speed and efficiency. Plus exciting new entrants from India and beyond.
72 Rubin GPUs per rack, NVLink-C2C fabric, NVIDIA Vera ARM host CPU. Powers "agentic AI factories" worldwide. Blackwell Ultra MLPerf v6.0 submission showed 2.77x speedup on DeepSeek-R1 over previous gen. Only vendor to submit DeepSeek-R1 results.
NVIDIA newsroom →Single chip the size of a dinner plate with ~21 PB/s on-chip memory bandwidth. Holds entire models in SRAM, eliminating memory bottleneck. 2,100 TPS on Llama 3.3 70B, 1,800 TPS on 8B. Verified by Artificial Analysis. 4-6x faster than Groq on identical models.
Cerebras blog →Compiler-driven architecture pre-computes entire execution graph. Near-zero latency variance. Best for voice assistants and real-time apps where predictable TTFT matters more than peak throughput. ~150ms p50 TTFT on Llama 3.3 70B.
groq.com →Reconfigurable Dataflow Unit built for agentic workloads. Three-tier memory architecture supporting models up to 10T parameters. Claims 895 TPS per user on Llama 3.3 70B with FP8, nearly 5x NVIDIA B200. 1.6 PFLOPS FP16.
SambaNova blog →Tensor Contraction Processor (TCP), a clean-sheet design vs GPU "legacy tax." RNGD chip in mass production at TSMC. 3rd-gen chiplet platform with Broadcom. SDK with general compiler maps PyTorch to silicon. CUDA-alternative with deterministic performance.
FuriosaAI blog →Kerala-based startup's flagship AI SoC for edge devices: smart cameras, automotive, edge AI boxes. Built on TSMC 12nm with in-house NPU, vision, and security engines. Rs 15 crore DLI scheme support. A4000 server chip expected Q2 2027.
ET report →Latency and throughput across 26+ providers. Measured, not marketed.
Wafer-scale engine eliminates memory bottleneck. ~520 tokens/s throughput. For real-time UI where sub-300ms TTFT is the difference between "instant" and "waiting." Groq follows at ~150ms p50 with ~750 tok/s throughput.
Cerebras benchmarks →10-20x cheaper than GPT-4o while still scoring 63/100 on quality. Perfect for high-volume classification, routing, and simple summarization. For 100k MAU chatbot: ~$107/month vs $3,575 on GPT-4o.
VerticalAPI benchmarks →The default recommendation for speed-sensitive production. Chatbot Arena Elo 1335. Fast enough for smooth streaming, cheap enough for scale, smart enough for most tasks. 25x cheaper than GPT-5 (high) with comparable quality for 90% of prompts.
Top-tier across coding, reasoning, and creative. But 5x the cost of Sonnet 4.5 (91/100) for only 3 quality points. Use Opus for the 5% of queries where marginal quality matters. Sonnet is the right default.
Diffusion-based architecture generates tokens in parallel, not sequentially. 13x faster than Claude 4.5 Haiku. Matches Haiku on reasoning benchmarks (AIME 2025: 91.1%). The speed advantage is architectural. It doesn't require custom silicon.
Mercury 2 review →GPT-4o swings from 820ms p50 to 1.9s p95. Cerebras and Groq have near-zero variance. For voice assistants and real-time chat, design around p95. That's what users actually feel. A single 8-second outlier ruins the experience.
Speed leaderboard →Sources: VerticalAPI · Awesome Agents · Artificial Analysis
Open models aren't catching up. They're leading. This is why AI freedom matters.
Dominates open-weight deployment. Qwen3-0.6B (19M downloads), Qwen3-8B (12M), Qwen3.5-Max (Elo 1465). Apache 2.0. From 0.6B to 397B MoE. The default for self-hosted agents.
R1 is the only open-weight reasoning model cited as frontier-grade. V4-Flash: $0.14/$0.28 per M. V4 Pro: Elo 1467, SWE-bench 80.6%. 75% price cut in June 2026. 1M context.
Unified multimodal, encoder-free. 12B variant runs on a single 4090. AIME 2026: 77.5. LiveCodeBench v6: 72.0. 256K context. The new local AI standard.
Frontier quality at $0.73/$3.49 per M. 12+ hour autonomous runs documented. Deep Research, Sheets, Agent Swarm, Kimi Code. One of the most included open-weight options.
Source: Presenc AI May 2026 Report