AI Model Releases, New LLMs, Papers, and Open Source AI: April 11-13, 2026

The three days after April 10 brought a different kind of AI progress than the big-name launches earlier in the week. Instead of headline model drops, April 11-13 delivered API rollouts for models announced days earlier, a surge of open source tooling, and research papers that quietly advanced the state of the art in reasoning and alignment. This post covers everything that shipped, published, or went live across all four categories.

At a Glance: What Happened April 11-13

| Date | Category | What | Why It Matters | |---|---|---|---| | Apr 11 | Model release | Meta Muse Spark API announced | First proprietary Meta model gets developer access | | Apr 11 | LLM update | GLM-5.1 GGUF quantizations hit Hugging Face | 754B MoE model now runnable locally via llama.cpp | | Apr 11 | Open source | Archon v2.1 ships harness builder | First open source AI coding harness builder, 14K stars | | Apr 11 | Open source | OpenAI Codex CLI gets Realtime V2 | Streaming audio and MCP support in terminal | | Apr 12 | Paper | AI Scientist-v2 conference acceptance | First fully AI-generated paper accepted at a major venue | | Apr 12 | Paper | PaperOrchestra multi-agent framework | 84% simulated acceptance rate at CVPR | | Apr 12 | LLM update | MiniMax M2.7 open sourced | Self-evolving agent model, $0.30/M input tokens | | Apr 12 | Open source | Ollama v0.20.6 bug-fix release | Stability fixes for Gemma 4 and GLM-5.1 backends | | Apr 13 | Paper | Sequence-Level PPO (SPPO) | New alignment technique combining PPO efficiency with outcome stability | | Apr 13 | Open source | claude-mem crosses 46K GitHub stars | Persistent memory plugin for Claude agents |

Model Releases and LLM Updates

Meta Muse Spark API (April 11)

Meta's first proprietary AI model, Muse Spark, debuted on April 8 inside WhatsApp, Instagram, Facebook, and Messenger. On April 11, Meta confirmed developer API access is coming. The model originates from Meta Superintelligence Labs (MSL), the division led by Yann LeCun and Alexandr Wang after their $14B partnership.

Key specs worth noting:

| Spec | Value | |---|---| | Architecture | Natively multimodal (text, image, voice) | | Context window | 262K tokens (Artificial Analysis), up to 1M (unconfirmed) | | MMMU-Pro | 80.5% | | Humanity's Last Exam | 39.9% (58% in Contemplating mode) | | HealthBench Hard | 42.8 (beats GPT-5.4's 40.1) | | Compute efficiency | 10x reduced vs. Llama 4 via "thought compression" |

Muse Spark marks a philosophical shift for Meta. After years of open sourcing Llama models under permissive licenses, MSL chose to keep Muse Spark proprietary. The API announcement on April 11 signals that Meta wants to compete directly with Anthropic and OpenAI on the commercial API front, not just the open source leaderboard.

GLM-5.1 Ecosystem Growth (April 11-12)

Zhipu AI's GLM-5.1 (released under MIT license on April 7) saw rapid community adoption over the weekend. The model is a 754B parameter Mixture-of-Experts architecture designed for agentic engineering tasks that can run autonomously for up to 8 hours.

Benchmark results that drove adoption:

| Benchmark | GLM-5.1 | Claude Opus 4.6 | Notes | |---|---|---|---| | SWE-Bench Pro | 58.4 | 57.3 | #1 open source | | Terminal-Bench | #1 open source | N/A | Agentic terminal tasks | | NL2Repo | #1 open source | N/A | Natural language to full repo | | LM Code Arena | 3rd overall | N/A | Behind two proprietary models |

The MIT license makes GLM-5.1 the most permissively licensed frontier-class model available. For teams running local inference, the new GGUF quantizations mean you can serve it through llama.cpp on high-end workstations.

MiniMax M2.7 Open Source (April 12)

MiniMax released M2.7 as an open source self-evolving agent model. Available through Together AI at $0.30 per million input tokens with a 205K context window. The "self-evolving" label refers to the model's ability to refine its own tool-use strategies during long agentic runs without human intervention.

Ongoing Rollouts from Earlier Launches

Several models announced before April 11 continued rolling out during this window:

• Claude Mythos Preview (announced April 7) remained in gated deployment through Project Glasswing, with 50 organizations running defensive cybersecurity scans. Anthropic committed $100M in usage credits and $4M in donations to open source security organizations.
• Google Gemma 4 (released April 2) saw critical GGUF quantization fixes land on April 12, resolving inference accuracy issues that affected the 26B MoE variant.
• Alibaba Qwen3.6-Plus (released April 2) continued gaining traction in agentic coding workflows, with 1M token context and compatibility with OpenClaw, Claude Code, and Cline.

Research Papers: April 11-13

AI Scientist-v2: Automated Scientific Discovery

The biggest paper news of the weekend was the confirmation that a research paper fully generated by AI Scientist-v2 was accepted at a major machine learning conference. The system uses agentic tree search to automate the scientific discovery pipeline, from hypothesis generation through experiment design, execution, and paper writing.

PaperOrchestra: Multi-Agent Paper Writing

Published alongside AI Scientist-v2's news, PaperOrchestra takes a different approach. Instead of end-to-end automation, it converts existing pre-writing materials (experiment logs, notes, data) into submission-ready manuscripts using coordinated agents. The team reported simulated acceptance rates of 84% at CVPR and 81% at ICLR.

Sequence-Level PPO (SPPO)

A new alignment technique combining PPO's sample efficiency with outcome-based reward stability. SPPO operates at the sequence level rather than token level, which reduces reward hacking on reasoning tasks. Early results show improved consistency on math and coding benchmarks compared to standard RLHF approaches.

In-Place Test-Time Training (In-Place TTT)

This paper repurposes MLP blocks for chunk-wise updates during inference, improving long-context performance without fine-tuning. On the RULER benchmark for long-context evaluation, In-Place TTT showed measurable gains for sequences beyond 128K tokens. The technique is model-agnostic and could be applied to any transformer architecture.

Open Source AI Projects

Archon: First Coding Harness Builder (14K+ Stars)

Archon (coleam00/Archon) shipped v2.1 during this window, establishing itself as the first open source harness builder for AI coding agents. It uses YAML-based workflow definitions with git worktree isolation, letting you define multi-step coding tasks that agents execute in sandboxed environments. At 14,000+ GitHub stars, it is growing fast.

OpenAI Codex CLI: Realtime V2 and MCP (5.8K Stars)

OpenAI's Codex CLI (openai/codex-cli) received streaming audio support via Realtime V2 and Model Context Protocol (MCP) integration. The update means you can talk to Codex in your terminal while it reads from external tools through MCP. Sandboxed execution remains the default, so generated code runs in isolation.

Other Notable Releases

| Project | Stars | What Shipped | |---|---|---| | google/adk-python | 8,200+ | Google Agent Development Kit for multi-agent systems | | meta-llama/llama-stack | 6,400+ | Unified deployment stack for Llama 4 family | | block/goose | 4,900+ | Local-first AI agent framework with MCP support | | claude-mem | 46,000+ | Persistent memory plugin for Claude agents | | Ollama v0.20.6 | N/A | Bug fixes for Gemma 4 and GLM-5.1 backend stability |

Industry Context: Why This Weekend Mattered

The April 11-13 period sits in a specific moment for the AI industry. OpenAI surpassed $25B in annualized revenue and is moving toward a public listing. Anthropic is approaching $19B annualized. A PwC study published on April 13 found that 75% of AI's economic gains are being captured by the top 20% of companies, and Gallup reported that 50% of employed Americans now use AI at work at least a few times per year.

Against that backdrop, the weekend's releases tell a story about distribution, not just capability. Meta is moving proprietary. Zhipu AI is going full MIT license. Research labs are building systems that write their own papers. And open source projects are racing to wire everything together.

Common Pitfalls When Tracking AI Releases

• Confusing announcement date with availability date. Muse Spark was "launched" April 8 but the API was not announced until April 11, and developer access has no confirmed date yet. Claude Mythos was announced April 7 but remains gated to 50 organizations.
• Assuming open weights means open source. Llama 4's community license has restrictions on commercial use above certain thresholds. GLM-5.1's MIT license and Gemma 4's Apache 2.0 license are genuinely permissive. The distinction matters if you are building a product.
• Ignoring quantization quality. Running a 754B MoE model via GGUF quantization on consumer hardware involves real accuracy tradeoffs. The Gemma 4 GGUF fixes on April 12 existed because early quantizations had measurable inference degradation.
• Treating benchmark numbers as absolute. GLM-5.1 beats Claude Opus 4.6 on SWE-Bench Pro (58.4 vs 57.3), but that is one benchmark measuring one type of task. Real-world performance depends on your specific workload.

Checklist: Staying Current After April 13

If you are building on top of these models and tools, here is what to watch next:

Wrapping Up

April 11-13, 2026 was not about single blockbuster launches. It was about the ecosystem absorbing and building on top of what dropped earlier in the week. Meta committed to a proprietary API, GLM-5.1 became practically usable through open source tooling, research papers showed AI systems writing their own papers, and the agent framework space continued its rapid expansion. The real story is velocity: the gap between model announcement and community integration is now measured in days, not months.

Fazm is an open source macOS AI agent. Open source on GitHub.