Open source LLM releases in May 2026: one new model so far, four April tail releases, and the Swift block that decides which of them drive your Mac

As of May 13, 2026, the only major new open-weight model released this month is OpenBMB MiniCPM-V 4.6 1.3B (released May 11, 2026, Apache 2.0). It is a 1.3B parameter multimodal vision-language model with a 262K token context, accepting text, image, and video inputs. The release sits inside a 2026 open-weight calendar where the dominant story is the four late-April drops still being benchmarked everywhere: Xiaomi MiMo-V2.5-Pro on April 28 (MIT, 1.02T total / 42B active MoE), NVIDIA Nemotron 3 Nano Omni on April 28 (NVIDIA Open Model Agreement, 30B / 3B active multimodal, runs in 25GB of RAM), IBM Granite 4.1 on April 29 (Apache 2.0, dense 3B / 8B / 30B with up to 512K context), and Mistral Medium 3.5 on April 29 (Modified MIT, 128B dense, 256K context). May is shaping up as a slow month relative to the late-April cluster.

Two reasons. First, the late-April week of April 28 to 29 absorbed most of the spring release energy, with four frontier-class labs (Xiaomi, NVIDIA, IBM, Mistral) shipping within 48 hours of each other; 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), which tends to soak up benchmark airtime even for labs that are not competing on the same axis. The MiniCPM-V 4.6 1.3B release on May 11 is a deliberate small-model play on a separate axis (multimodal SLM that fits on a phone), so it is not really competing with the late-April large drops for headline space. Several open-weight labs are publicly aiming at end of May for their next drop. There is no fundamental reason May ends up thin; the calendar mechanics just landed this way.

The model weights only get you halfway. You also need a consumer Mac agent that will let you point its inference target at a proxy server instead of the default Anthropic API. In Fazm specifically that is a single Settings toggle backed by a three-line block at lines 522 to 525 of Desktop/Sources/Chat/ACPBridge.swift. The block reads the customApiEndpoint string from UserDefaults and assigns it to env["ANTHROPIC_BASE_URL"] on the spawned ACP subprocess. With that env var pointed at a LiteLLM, vLLM, or any Anthropic-compatible proxy that re-emits the request as MiniCPM-V 4.6, MiMo-V2.5-Pro, Nemotron 3 Nano Omni, Granite 4.1, or Mistral Medium 3.5, the agent runs end to end against your real macOS accessibility tree. The code is visible to anyone reading the open repo at github.com/m13v/fazm.

Three of the five are genuinely permissive. MiniCPM-V 4.6 1.3B and IBM Granite 4.1 ship under Apache 2.0 with no field-of-use or revenue clauses. Xiaomi MiMo-V2.5-Pro ships under MIT, also unconditional. Mistral Medium 3.5 ships under a Modified MIT license that is permissive for individuals, startups, mid-market companies, and most universities, but requires a negotiated commercial agreement for companies above a revenue threshold; if you are deploying inside a large company you have to read the actual license terms. NVIDIA Nemotron 3 Nano Omni ships under the NVIDIA Open Model Agreement, which permits commercial use but is not OSI-approved and has its own redistribution and attribution rules. If you need a clean, unencumbered weights file for a small business workflow on a Mac, the Apache 2.0 and MIT entries are the safe picks.

For most agent workloads no, for two specific workloads yes. The model is optimized for vision-language efficiency on a phone (LLaVA-UHD v4 backbone, roughly 50% fewer visual encoding FLOPs than the prior version, about 1.5x token throughput versus Qwen3.5-0.8B). For the general desktop agent loop, where the model needs to plan multi-tool sequences and reason over the macOS accessibility tree, a 1.3B parameter model is undersized; the late-April drops are the realistic targets there. Where MiniCPM-V 4.6 is genuinely interesting on a Mac is for the always-on screen understanding loop (reading what is currently visible without spinning up a frontier model for each frame) and for any flow that has to fall back to an offline pass when the network is unreliable. Those are exactly the cases where a 1.3B model with a 262K context is a better fit than a 100B-class frontier model.

Most teams run an Anthropic-compatible proxy in front of the actual inference server. LiteLLM has shipped Anthropic API translation for several of these models out of the box and is the most common bridge for MiMo-V2.5-Pro, Granite 4.1, and Mistral Medium 3.5. vLLM 0.20.1 (the May 3, 2026 release) 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. For Nemotron 3 Nano Omni, NVIDIA's NIM microservice is the official path; you wrap it in LiteLLM the same way. In all cases the only thing the Mac agent needs to know is the proxy URL; the three-line block in ACPBridge.swift does the rest.

Yes, and they matter for the agent layer. Hugging Face shipped ml-intern (the open-source post-training automation agent built on the smolagents framework) on April 21, 2026, which is open code but not a model release. Several inference servers also shipped May 2026 versions that are open source but not models: vLLM v0.20.1 on May 3 stabilizes DeepSeek V4 inference and ships a configurable VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD knob. These are part of the May 2026 open-source picture even though they are not LLM weight releases.

The four-release cluster on April 28 and 29 is genuinely load-bearing for the May 2026 open-source LLM conversation, and any roundup that omits them is misleading. The shortest honest summary: Xiaomi MiMo-V2.5-Pro is the largest by raw parameter count (1.02T MoE, 42B active) and shipped under MIT, which makes it the most permissive frontier-class release. NVIDIA Nemotron 3 Nano Omni is the efficiency story (30B total / 3B active multimodal, runs in 25GB of RAM, claimed up to 9x higher throughput than other open omni models on video and document workloads). IBM Granite 4.1 is the practical enterprise pick (Apache 2.0 dense 3B/8B/30B with up to 512K context and 12 supported languages). Mistral Medium 3.5 is the 128B dense coder (256K context, 77.6% on SWE-Bench Verified) under a Modified MIT license that you need to read carefully for commercial deployments.

Probably yes. The pattern through 2026 has been one to three open-weight releases of any size per week, and the May 1 to May 13 window has only produced MiniCPM-V 4.6 1.3B. The labs publicly aiming at late May include the GLM team (GLM-5.2 has been previewed) and at least two China-based labs that signaled May drops without dates. The list on this page reflects what is verifiably shipped as of May 13, 2026; it will be updated as actual releases land.