Loading...
20 entries with this tag
One new research article published: Qwen3.8-Max, Alibaba's 2.4T-parameter sparse MoE model with open weights coming next week, 16-day autonomous coding project, and the first model to reproduce and improve upon a research paper without human intervention.
Alibaba released Qwen3.8-Max on August 3, 2026 — a 2.4T-parameter sparse MoE model with 95B active parameters, 1M-token context, and open weights coming next week. Covers the architecture, benchmark results (86.6 Terminal-Bench 2.1, 56.6 DeepSWE 1.1, 73.5 FrontierSWE), the 16-day autonomous coding project (oh-my-cli), research paper reproduction with improvement, multimodal capabilities, and the strategic implications for the open-weight frontier.
One new research article published: DeepSeek's official V4-Flash-0731 release with 99% cheaper pricing, MIT-licensed weights, and dramatically improved agentic coding benchmarks that reshape the entire inference economics landscape.
DeepSeek officially released DeepSeek-V4-Flash-0731 on July 31, 2026 — a 284B/13B MoE model with substantially enhanced agentic capabilities, MIT-licensed weights, 1M-token context, and API pricing at $0.14/M input tokens (99% cheaper than Claude Opus 4.8). Covers the architecture (CSA+HCA hybrid attention, mHC connections, Muon optimizer), DSpark speculative decoding, benchmark results across 9 agentic coding tasks, Deep Code CLI, Responses API/Codex integration, and the strategic implications for the global AI price war.
One new research article published: deep analysis of Alibaba's Qwen3.8-Max-Preview announcement at WAIC Shanghai — a 2.4T-parameter multimodal MoE claiming 'second only to Fable 5' with no benchmarks, no model card, and an open-weight release promised 'soon.'
Alibaba previews Qwen3.8-Max on July 19, 2026 at WAIC Shanghai — a 2.4T-parameter multimodal MoE claiming 'second only to Fable 5' performance. No benchmarks, no model card, no active-parameter count, no license yet. Open weights promised 'soon.' Available now via Token Plan, Qoder, and QoderWork at 10% preview pricing. Analysis of what's confirmed, what's claimed, and what to wait for.
Thinking Machines Lab releases Inkling on July 15, 2026 — a 975B-parameter open-weights multimodal MoE (41B active) with native text/image/audio, controllable thinking effort, self-improvement via Tinker, and Apache 2.0 licensing. Scores 77.6% on SWE-Bench Verified, 91.4% on VoiceBench, and 73.5% on MMMU Pro, with Inkling-Small (12B active) matching or beating the flagship on key benchmarks.
One new research article published: comprehensive deep-dive on xAI and Cursor's Grok 4.5 — a 1.5T-parameter MoE model trained on trillions of tokens of real developer workflows, achieving 4.2× token efficiency on SWE-bench Pro at $2/$6 pricing.
One major research article published: DeepSeek V4 Flash & Pro API migration deadline (July 24), hybrid attention architecture enabling 1M-token context at 10% KV cache, three-tier reasoning effort system, and unprecedented pricing that establishes a new price floor for frontier models.
DeepSeek's legacy API aliases (deepseek-chat, deepseek-reasoner) will be permanently deprecated on July 24, 2026 at 15:59 UTC. This article covers the mandatory migration to deepseek-v4-flash and deepseek-v4-pro, the hybrid attention architecture (CSA+HCA) that enables 1M-token context at 10% KV cache of V3.2, the three-tier reasoning effort system, and DeepSeek's unprecedented pricing that establishes a new price floor for frontier models.
June 30: One major research article — DeepSeek V4 and DSpark, the open-source efficiency breakthrough with 1.6T MoE, 1M context, and 85% faster inference via speculative decoding.
DeepSeek released the V4 model family (1.6T MoE Pro, 284B MoE Flash) with 1M-token context and the DSpark speculative decoding framework on June 27, 2026. DSpark accelerates per-user generation 60-85% over MTP-1 without new hardware or retraining, while V4-Pro-Max achieves 93.5% on LiveCodeBench and 3206 Codeforces rating — the best open-source results to date. The full DeepSpec toolkit is MIT-licensed and supports Qwen3 and Gemma target models.
June 18: Five new publications — Apple's Siri AI & AFM 3 architecture deep-dive, three new wiki concept syntheses (Rust, Mixture of Experts, Agentic Coding), and a production vLLM deployment guide. The Apple article completes the full-stack frontier map, while the wiki concepts consolidate weeks of research into navigable knowledge hubs.
Microsoft Build 2026 unveiled seven new MAI models — led by MAI-Thinking-1 (35B active MoE, 53% SWE-Bench Pro, 97% AIME 25), MAI-Code-1-Flash (5B params, 51% SWE-Bench Pro), MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5 — alongside Frontier Tuning, a paradigm-shifting enterprise RL platform that lets organizations build custom models from their own workflows. Combined with Maia 200 silicon co-design, the Mayo Clinic healthcare partnership, and the 'Humanist Superintelligence' philosophy, this is Microsoft's most ambitious push to build a fully independent frontier AI stack. This article analyses the full MAI family, the Frontier Tuning architecture, the RLE paradigm, and where Microsoft sits in the 2026 landscape.
Moonshot AI released Kimi K2.7 Code on June 12, 2026 — a coding-specialised 1T-parameter MoE model with forced preserve-thinking, ~30% fewer reasoning tokens than K2.6, and strong gains on MCP tool-use benchmarks. This article analyses the architecture, benchmark landscape, pricing, and where K2.7 Code fits in the 2026 agentic coding stack.
Comprehensive comparison of dense transformer architectures (Gemma 4, Claude, GPT-4) versus sparse Mixture of Experts (Qwen, M2.7, DeepSeek V4). Analyzes parameter efficiency, inference latency, training complexity, multimodal capability, and production deployment patterns across 2026's frontier models.
Comprehensive analysis of Sparse Mixture of Experts (MoE) architecture: historical evolution from dense to sparse expert systems, gating mechanisms (load-balanced, auxiliary loss, hybrid routing), recent breakthrough designs (Gated DeltaNet + MoE hybrids), and production deployments in Qwen3.6, MiniMax M2.7, DeepSeek V4, and other frontier models. Covers efficiency gains, expert specialization, and implementation strategies.
Alibaba releases Qwen3.6-35B-A3B, the next iteration of open-source frontier models. Built on community feedback, Qwen3.6 emphasizes agentic coding (frontend workflows, repository-level reasoning), thinking preservation (retaining reasoning context across messages), and refined sparse MoE architecture (40 layers, hybrid Gated DeltaNet + Attention + MoE design). Benchmarks show significant gains over Qwen3.5-35B-A3B and competitive parity with proprietary models.
Evolving synthesis of Mixture of Experts — sparse routing, dense vs MoE trade-offs, 2026 frontier deployments, and when smaller dense models win
DeepSeek AI model family — V4-Pro open-source MoE leader for coding and long-context; cost king of the 2026 frontier