Large language model releases in May 2026

Verified as of May 29, 2026, the month's notable releases split into proprietary and open weight. Proprietary: OpenAI made GPT-5.5 Instant the default ChatGPT model on May 5; xAI's Grok 4.3 reached general availability in early May (its API model-retirement notice went out May 6); Google made Gemini 3.5 Flash generally available around May 19 at Google I/O; and Anthropic shipped Claude Opus 4.8 on May 28. Open weight, both on Hugging Face: OpenBMB MiniCPM-V 4.6 1.3B on May 11 (Apache 2.0, 262K context), and Cohere Command A+ on May 20 (Apache 2.0, 218B total / 25B active MoE, 128K input). The wider conversation is still carried by four frontier-class open-weight models that landed within 48 hours of each other on April 28 and 29 (Xiaomi MiMo-V2.5-Pro, NVIDIA Nemotron 3 Nano Omni, IBM Granite 4.1, Mistral Medium 3.5), which is why an honest May roundup includes them.

No. GPT-5.5 Instant, which OpenAI made the ChatGPT default on May 5, 2026, is a proprietary model with no weights file. For a Mac agent there are two practical paths. Fazm bundles a Codex backend (the CodexBackendManager in Desktop/Sources/CodexBackendManager.swift, behind an enableCodexBackend toggle) that runs GPT models through your existing ChatGPT subscription via OpenAI's Codex backend; its model picker lists gpt-5.5 variants directly. The other path, for any proprietary model, is an Anthropic-compatible gateway plugged into the custom endpoint field. Either way you reach the model without a weights file.

Yes, through one of three paths, all visible in the open repo at github.com/m13v/fazm. For Claude (including Opus 4.8 from May 28) you bring your own Claude Pro or Max account: the bridge runs in personalOAuth mode (BridgeMode.personalOAuth in Desktop/Sources/Chat/ACPBridge.swift) and usage hits your existing plan. For GPT-5.5 you use the bundled Codex backend and your ChatGPT subscription. For anything else (Grok 4.3, Gemini 3.5 Flash, or an open weight) you point the customApiEndpoint setting at an Anthropic-compatible gateway, which the bridge assigns to ANTHROPIC_BASE_URL on the spawned agent subprocess. The same agent loop then drives your real macOS accessibility tree.

Almost every May 2026 release leads with a bigger context window. That number only helps if the tool you run the model through does not silently discard your conversation when it fills up. The raw Claude Code CLI auto-compacts: as the live context nears its limit it summarizes earlier turns and drops the originals, which is exactly when a long agent run starts losing the decisions you made an hour ago. Fazm does not auto-compact; the full chat history stays live in context for the lifetime of the window. So when you point it at Command A+ at 128K input or Granite 4.1 at up to 512K, the window can actually use that headroom.

Because the four-release cluster on April 28 and 29, 2026 is genuinely load-bearing for the May conversation. Xiaomi MiMo-V2.5-Pro (MIT, 1.02T total / 42B active MoE) is the largest by raw parameter count and the most permissive frontier-class license. NVIDIA Nemotron 3 Nano Omni (30B / 3B active multimodal, runs in 25GB of RAM) is the efficiency story. IBM Granite 4.1 (Apache 2.0, dense 3B / 8B / 30B, up to 512K context) is the practical enterprise pick. Mistral Medium 3.5 (Modified MIT, 128B dense, 256K context, 77.6% on SWE-Bench Verified) is the EU-friendly coder. As of late May these four are still the bar every new release is measured against.

Of the two open-weight May releases, Cohere Command A+ is the credible agent-planning target and MiniCPM-V 4.6 1.3B is not. Command A+ is a 218B / 25B-active Sparse MoE with 128K input context under Apache 2.0, shipped on Hugging Face in BF16, FP8, and W4A4 quantizations on day one, so it can run on as few as two H100s. MiniCPM-V 4.6 1.3B is a 1.3B multimodal model optimized for on-device inference; on a Mac it is a fit for an always-on screen-understanding pass, not the multi-step planning loop. For planning, the late-April frontier drops (MiMo-V2.5-Pro, Granite 4.1, Mistral Medium 3.5) are the realistic open-weight targets.

Fazm forks a conversation in one step. The forkSession(fromKey:toKey:cwd:model:) call in Desktop/Sources/Chat/ACPBridge.swift branches the active session into a new window, leaves the source session intact and resumable, and lets the branch carry a different model than the source. So when a new model ships you can branch a real working thread, switch the branch to the new model (or a different backend), and run the same prompt, without rebuilding context or touching the original.

Yes. Fazm persists chat history to a local SQLite database (the ChatMessageStore layer in Desktop/Sources/ChatMessageStore.swift) and reloads the session on launch, so windows come back with their history intact. This matters during a release cycle specifically: comparing models is a multi-day activity, and a restart should not wipe the threads you set up to evaluate GPT-5.5, Grok 4.3, Command A+, or Opus 4.8.

No. The pattern across 2026 has been one to three open-weight releases of any size per week, plus proprietary frontier drops on their own cadence (Claude Opus 4.8 landed as late as May 28). The list on this page is a snapshot verified as of May 29, 2026. Treat it as a moving picture, not a final accounting of the month.

The open-weight models have public pages: huggingface.co/CohereLabs/command-a-plus-05-2026-bf16 for Command A+, huggingface.co/openbmb/MiniCPM-V-4.6 for MiniCPM-V 4.6, huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro for MiMo-V2.5-Pro, and research.ibm.com/blog/granite-4-1-ai-foundation-models for Granite 4.1. The proprietary releases are documented on the OpenAI, xAI, Google, and Anthropic announcement channels (for example anthropic.com/news for Claude Opus 4.8 and openai.com for GPT-5.5 Instant). The Fazm behaviors described here are readable in the open repo at github.com/m13v/fazm.