Open-source AI projects, tools, and updates in May 2026: the agent-layer month

The verifiable May 2026 ledger has four layers. Open-weight LLMs on Hugging Face: Cohere Command A+ on May 20 (Apache 2.0, 218B sparse MoE, 128K context) and OpenBMB MiniCPM-V 4.6 on May 11 (Apache 2.0, 1.3B multimodal, 262K context). Inference runtimes: vLLM v0.21.1rc0 on May 15 (KV offload, Blackwell backend) and MLX Engine v1.8.1 on May 13 (Apple Silicon parallel predictions). Agent and IDE tooling: LangChain 1.3.0/1.3.1/1.3.2 across the month, LlamaIndex 0.14.22 on May 14, Claude Code v2.1.152 on May 27 (/code-review --fix), OpenAI Codex CLI 0.129 and 0.130 (vim mode, Goal Mode), Cline v3.82 (May 2) and v3.83 (May 14). Infrastructure: EleutherAI lm-evaluation-harness v0.4.12, the MCP 2026-07-28 spec release candidate locked on May 21, Argmax OSS SDK 1.0 graduating from WhisperKit. And at the consumer-agent layer, the open-source macOS agent Fazm shipped 43 versions between May 1 and May 28, every release stamped in CHANGELOG.json at github.com/mediar-ai/fazm.

Two reasons. First, the late-April week of April 28 to 29 absorbed most of the spring release energy: Xiaomi MiMo-V2.5-Pro, NVIDIA Nemotron 3 Nano Omni, IBM Granite 4.1, and Mistral Medium 3.5 all shipped within 48 hours. That is a hard release pattern to follow inside the same calendar month. Second, the proprietary side took the May 5 to May 6 window (OpenAI GPT-5.5 Instant on May 5, xAI Grok 4.3 on May 6) and Google opened I/O 2026 with Gemini 3.5 Flash on May 19, which soaked up benchmark airtime. Two new permissive open-weight releases (Cohere Command A+ and MiniCPM-V 4.6) is honest for a month following four frontier-class April drops.

Yes. It is the most credible May 2026 Hugging Face release for the agent planning loop, where MiniCPM-V 4.6 1.3B is not. Command A+ is a 218B parameter sparse Mixture-of-Experts with 25B active per token, 128K input context, under a clean Apache 2.0 license. Cohere shipped it on Hugging Face in BF16, FP8, and W4A4 quantizations on day one, so it can run on as few as two H100s (or a single Blackwell at 4-bit). The Apache 2.0 license means there is no field-of-use clause to clear before using it in a small business workflow. It is Cohere's first fully OSS model.

Five things worth tracking. (1) A stateless protocol core, so a server no longer has to hold per-client state to be spec-compliant. (2) An Extensions framework that lets a server advertise capabilities outside the core surface without forking the spec. (3) Tasks, a first-class long-running operation abstraction with explicit progress events. (4) MCP Apps, a packaging shape for distributing a server plus its UI hints together. (5) OAuth and OIDC-aligned authorization, replacing the looser shapes the spec previously left to implementers. The streamable-HTTP transport now requires Mcp-Method and Mcp-Name headers, and list and read results carry ttlMs and cacheScope fields so clients can cache deterministically. The RC was locked May 21, 2026 with a final release targeted for July 28.

Fazm v2.7.1 on May 1, with 29 changes in a single CHANGELOG.json object. The headline items were the experimental Codex backend (Settings, Advanced, Codex Backend) that routes GPT-5 family models through OpenAI's Codex CLI using a personal ChatGPT account; Heal cards on the Discovered Tasks surface that flag Mac health symptoms like kernel panics; the switch to a flat $9.99 per month subscription with no free trial and paywall at launch; the sticky stack of recent user messages at the top of every chat window so prompts stay visible while you scroll deeper; and the fix for orphan ACP bridge processes accumulating across crashes (each one had been holding ~600MB of RAM). v2.9.41 on May 27 came close with 12 changes including the verification-codes skill that pulls 2FA codes from Messages, WhatsApp, or Notification Center instead of asking the user to read them out.

43 releases between v2.7.1 (May 1) and v2.9.47 (May 28), 171 individual change strings in the changes arrays, averaging 15.1 hours between releases over the 27-day window. The cadence was not uniform: the largest single day was May 27 with six releases (v2.9.41 through v2.9.46) consolidating chat-window scroll and CPU fixes. The full record is in CHANGELOG.json at the root of the repo, mirrored at github.com/mediar-ai/fazm/blob/main/CHANGELOG.json. Each release object has version, date, and changes fields; you can grep '2026-05' on the file and count the matches yourself.

Three patterns that show up across the projects, not just Fazm. First, multi-backend swappability is now table-stakes for an agent: Fazm added Gemini Pro and Flash alongside Claude and ChatGPT in v2.9.35, the Codex CLI shipped per-server MCP env plus OAuth, Claude Code shipped /code-review --fix that operates on whatever the active model is. Second, the MCP layer is being treated as the integration substrate rather than a competing tool layer; the spec RC, the Codex per-server MCP work, the Fazm Assrt MCP for browser automation, the lm-eval-harness LiteLLM backend, all point at the same convergence. Third, the boring stability work (scroll handling, CPU profiling, recovery loops, paywall enforcement) is what occupies most of the surface area of a real production agent's monthly diff. The model-weight headlines drive the discovery; the changelogs are where the product actually gets useful.

Stand up an Anthropic-compatible proxy in front of the actual inference server. LiteLLM is the most common bridge for Command A+, MiMo-V2.5-Pro, Granite 4.1, and Mistral Medium 3.5; it has shipped Anthropic API translation for several of these out of the box. vLLM 0.21.x ships an OpenAI-compatible server and pairs with LiteLLM to expose an Anthropic surface. For MiniCPM-V 4.6 the easiest path is the OpenBMB serving repo plus a LiteLLM bridge. Once the proxy URL is set, point a wrapper agent's custom API endpoint at it. In Fazm specifically that is a single Settings toggle (Settings, Advanced, AI Chat, Custom API Endpoint) that is read into env['ANTHROPIC_BASE_URL'] on the spawned ACP subprocess. Same pattern works for any agent that respects the env var.

The notable open-source voice release was Argmax OSS SDK 1.0, which graduates the WhisperKit codebase into a broader SDK with WhisperKit Local Server (an OpenAI-compatible Vapor HTTP server) and Swift 6 concurrency. No new Moshi or Kyutai release shipped in May based on the primary sources we checked. On the consumer side, Fazm's voice work this month landed inside push-to-talk: v2.9.16 fixed the mic and PTT button needing two clicks in unfocused pop-out windows, v2.9.25 fixed PTT inserting repeated gibberish (such as 'भाई भाई भाई') when transcription hallucinated on silent audio, and v2.9.25 also fixed PTT transcription concatenating onto previously spoken text so each voice message starts fresh.

Three steps. (1) Open https://github.com/mediar-ai/fazm/blob/main/CHANGELOG.json in a browser, or clone the repo and open CHANGELOG.json at the root. (2) The file is a JSON object with two keys, unreleased and releases. The releases array is ordered newest to oldest; each entry has version, date, and changes. (3) Filter the array on date starting with '2026-05' and count the entries (43) and the total length of the changes arrays (171). The first May entry is v2.7.1 dated 2026-05-01, the last is v2.9.47 dated 2026-05-28. Versions v2.9.7, v2.9.11, v2.9.13, v2.9.14, v2.9.23, v2.9.24, v2.9.27, v2.9.28, v2.9.29, v2.9.38, v2.9.39, and v2.9.40 do not appear as separate dated entries; the work landed batched into adjacent versions.