Latest open source LLM releases, 2026

As of June 20, 2026 the most recent open-weight drops are MiniMax M3 (June 1, the first open-weight model to combine frontier coding, a 1M-token context, and native multimodality, with a vendor-reported 59.0% on SWE-Bench Pro), and then a mid-June pair: Kimi K2.7 Code from Moonshot AI and GLM-5.2 from Z.ai, both shipped the week of June 12 to 13. Behind those, the spring of 2026 produced DeepSeek V4 (previewed April 24, MoE in 1.6T and 284B sizes, MIT), Kimi K2.6 (April 20, 1T total / 32B active, Modified MIT), Mistral Medium 3.5 (April 29, 128B dense, Modified MIT), Google Gemma 4 (April, Apache 2.0), Qwen 3.5 (February, Apache 2.0), and Llama 4 Scout (Meta custom license). Apache 2.0 and MIT dominate the permissive end of the list; the Chinese frontier labs lean on Modified MIT with a revenue or attribution clause. This list is verified against vendor model cards and updated as new weights land.

DeepSeek V4 (MIT), Qwen 3.5 (Apache 2.0), and Google Gemma 4 (Apache 2.0) are the unconditional picks. MIT and Apache 2.0 carry no field-of-use, revenue, or monthly-active-user clauses, so you can run them inside a paid product without clearing terms. The Modified MIT licenses on the Kimi K2 line and Mistral Medium 3.5 are permissive for individuals and most companies but add a clause above a revenue or user threshold (Kimi requires a 'Kimi K2' UI attribution above 100M MAU or $20M monthly revenue). Llama 4 Scout uses Meta's custom license with a 700M MAU ceiling and is not OSI-approved. MiniMax M3's license has commercial conditions you should read before shipping. If you need a clean weights file for a small business workflow, start with the Apache 2.0 and MIT entries.

The weights only get you halfway. You also need a desktop agent that lets you redirect its inference target from the default Anthropic API to a proxy that re-emits requests as your chosen model. In Fazm that is a single Settings field, customApiEndpoint, backed by a block in Desktop/Sources/Chat/ACPBridge.swift (around lines 2417 to 2439). The block trims the value, validates that it is an absolute http or https URL with a host, sets env["ANTHROPIC_BASE_URL"] on the spawned agent subprocess, and swaps Fazm's bundled Anthropic key for a placeholder so your proxy never receives the real key. Point that field at a LiteLLM or vLLM gateway re-emitting DeepSeek V4, Qwen 3.5, Gemma 4, or any other model on this page, and the agent runs end to end against your real macOS accessibility tree. The code is visible in the open repo at github.com/m13v/fazm.

Because a malformed value silently bricks chat. If you paste 'localhost:8766' (no scheme) or stray text into the field and the app forwards it raw, the Anthropic SDK throws 'Invalid URL' on every query, and the retry-with-resume path can swallow that into an empty turn so you see no error at all. The validCustomAPIEndpoint helper (ACPBridge.swift, around lines 301 to 312) guards against this: it requires a non-empty trimmed string, a parseable URL, an http or https scheme, and a non-empty host. Anything else is ignored and the app falls back to the default Anthropic endpoint with a log line, so a typo can never leave you with a dead chat window.

Not all of them. MiniMax M3's launch figures (59.0% SWE-Bench Pro, 66.0% Terminal-Bench 2.1) were run on MiniMax's own infrastructure with agent scaffolding, and at launch the weights were not yet on Hugging Face and the parameter count was undisclosed, so treat those as vendor-reported until third-party reruns land. The same caution applies to GLM-5.1's 58.4% SWE-Bench Pro and other day-one numbers from the Chinese frontier labs. The ship dates and licenses on this page are the durable facts; the benchmark deltas move as independent evaluations come in.

Either works, because the override lives at the chat-engine layer, not in your model files. The customApiEndpoint field points the agent at whatever your proxy exposes, so the practical pattern is to run a gateway (LiteLLM is the most common) that routes by model name to several backends, then keep the Fazm endpoint pointed at the gateway. You change which weights answer by changing the model the gateway serves, not by reconfiguring the Mac app. For a small business workflow that mostly needs reliable tool calls over the accessibility tree, a clean Apache 2.0 model like Gemma 4 or Qwen 3.5 is a sensible default; reach for a 1M-context model like DeepSeek V4 when a single run has to hold a large document or a long browser session.

Yes, and it is the main reason to do it. When the custom endpoint is set, Fazm disables its bundled Anthropic key and substitutes a harmless placeholder (sk-fazm-custom-endpoint), so the real key never reaches your proxy and the request goes to the address you control rather than to Anthropic. If that proxy fronts a model you are hosting locally or in your own cloud, your prompts, your screen context, and your accessibility-tree reads stay inside infrastructure you own. That matters more for some 2026 releases than others: a model served from a third-party API in a jurisdiction with broad government-access laws is a different trust posture than the same open weights running on your own box.

Any Anthropic-compatible endpoint works, which includes proxies in front of open weights, corporate gateways, and GitHub Copilot-style routes. The field is named for custom endpoints precisely because it does not care what is behind the URL, only that the URL speaks the Anthropic API surface. That is what makes the open-weight calendar on this page usable from a Mac: the bridge code never special-cases a vendor, so the day a new permissive model ships and someone stands up an Anthropic shim for it, you can point Fazm at it without an app update.

It is open source. The endpoint forwarding lives in Desktop/Sources/Chat/ACPBridge.swift (the env["ANTHROPIC_BASE_URL"] assignment around lines 2417 to 2439), the validation helper is validCustomAPIEndpoint in the same file (around lines 301 to 312), and the Settings text field that writes the value is in Desktop/Sources/MainWindow/Pages/SettingsPage.swift with a placeholder showing the host and port shape of a local proxy (for example a host on port 8766). The whole repo is at github.com/m13v/fazm, so you can read the path a request actually takes before you trust it with your screen and keyboard.