Every model released this weekend, three labels a Mac app actually ships
April 11-12, 2026 was a heavy two-day window. Claude Mythos 5 hit ten trillion parameters. Gemma 4 landed under a cleaner Apache 2.0 license. Gemini 3.1 Ultra opened a 2-million-token native multimodal context. Google's TurboQuant paper made long context cheaper at ICLR. Every AI-news roundup stops there. This guide keeps going. While the big labs shipped papers and weights, one Mac app shipped a version on each of those same two days, and its source has the cleanest answer to the only question the reader actually cares about, which one do I pick, and can I route to the open-source model that just dropped.
Models, papers, and license events, April 11-12, 2026
What actually dropped the weekend of April 11-12, 2026
The big buckets the major AI-news sites agreed on. Four of them are model releases or paper releases, two of them are licensing or aggregate counts. All of them matter for what a Mac user can actually run on Monday morning.
Claude Mythos 5 (Anthropic)
First widely recognized 10-trillion-parameter model. Positioned for long-horizon agent workflows. Hosted only for now, no open weights.
Gemma 4 (Google, Apache 2.0)
Open-weight release, 9B and 27B variants. Updated commercial license on April 11 removed the old user-count cap. Lands in Ollama and MLX within days.
Gemini 3.1 Ultra (Google DeepMind)
Native multimodal: text, audio, image, video in one training objective. 2-million token context window. No transcription intermediary.
TurboQuant paper (ICLR 2026)
PolarQuant rotation plus Quantized Johnson-Lindenstrauss compression. Cuts KV cache memory for long-context serving. Economics shift, not a shipped model.
Gemma 3 license refresh (April 11)
Google dropped the prior user-count restriction on Gemma 3's commercial license the same weekend Gemma 4 dropped. Lowers friction for commercial reuse.
Seven open-source drops, first 12 days
Per LLM-Stats: seven major open-source model releases in the first twelve days of April 2026. Most of them landed in Ollama and MLX before the weekend.
The part every other roundup stops short of
Read any AI-news aggregator on the weekend of April 11-12 and the shape is the same. Here is Claude Mythos 5. Here is Gemma 4. Here is Gemini 3.1 Ultra. Here is the TurboQuant paper. Here are the seven open-source models that dropped in the first twelve days. End of article.
None of them answer the next question. Fine, these all exist, which one do I actually pick, and if the model that matters to me is one of those open-source drops, how do I use it from a real Mac app today, without writing any code?
That answer lives in a consumer Mac app that shipped two versions on those exact same dates. Version 2.2.0 on April 11 and version 2.2.1 on April 12. The commit that matters is not a new model, it is a setting: Custom API Endpoint. The screen that matters is one struct in one Swift file with three string literals in an array. The whole rest of this guide is about that struct and that setting.
What shipped in Fazm on April 11 and April 12, 2026
Two numbered releases back-to-back. Every entry below is copied directly from the CHANGELOG.json in the Fazm repo at the paths shown. No paraphrase, no marketing rewrite.
The new-feature line that matters for the rest of this guide is the first one in v2.2.0. Custom API endpoint setting for proxies and corporate gateways. That one line is the bridge from Claude-hosted inference to any open-source model that dropped the same weekend.
The three model labels a Mac app actually committed to
The entire model selector in Fazm's floating bar, the thing a user sees when they tap the model name, is driven by one static array of three elements. It is at lines 150 through 155 of Desktop/Sources/FloatingControlBar/ShortcutSettings.swift in the Fazm repo. It looks like this, verbatim.
Scary, Fast, Smart. Three labels, one file, nothing else. No feature flag, no A/B test, no remote config. The whole April 2026 buffet of model choice, at least for the default Fazm user, collapses to one of these three strings. If ACP reports a new model at runtime, the mapping at lines 157 to 162 below auto-renames it based on which substring it contains.
Every April 11-12 model lines up behind one Custom Endpoint field
On the left, every model or paper-adjacent release from the weekend that a user might actually want to try. In the middle, the Fazm custom-endpoint plumbing that normalizes them into one Anthropic-shaped request. On the right, a single Mac chat that drives any app open on your desktop.
From April 11-12 releases to one Mac chat
“On the weekend ten-trillion-parameter models and 2-million-token contexts shipped, the consumer-app question was answered by one text field, one env var, and three string literals in a Swift array.”
Fazm, CHANGELOG.json v2.2.0 (2026-04-11) and ShortcutSettings.swift:152-154
The Opus migration that quietly says Smart is not default
Fazm has three labels but does not treat them equally. Buried in the same ShortcutSettings.swift file, lines 287 to 296, there is a one-time migration that rewrites the stored model selection from Opus to Sonnet for anyone who had Opus saved. The comment is the whole story.
The takeaway for a reader navigating April 11-12 releases: a consumer app that runs real agent loops, with tool calls, retries, and multi-turn UI automation, deliberately downgrades users from the strongest model in its own picker. Not because Opus is bad at the task, because Opus burns rate limits too fast for the loops. If you are picking a model this weekend based on benchmark posts alone, that is the trade-off that benchmark posts never surface.
The two lines that route Fazm to any open-source model you want
The Custom API Endpoint feature added in v2.2.0 on April 11, 2026 is implemented in two lines. File is Desktop/Sources/Chat/ACPBridge.swift. The lines are 379 and 380. Before the ACP subprocess that talks to the model is spawned, Fazm reads the user's saved endpoint from UserDefaults and sets ANTHROPIC_BASE_URL on the subprocess env.
That is the entire redirection. Any inference server that speaks the Anthropic Messages API shape, a local Ollama proxy for Gemma 4, a vLLM bridge for DeepSeek V4, a corporate gateway for GitHub Copilot, a self-hosted endpoint for a Qwen 3.6 fine-tune, becomes a drop-in. No new Fazm release needed for any model that drops tomorrow.
The UI that writes that UserDefault is a TextField around line 933 of Desktop/Sources/MainWindow/Pages/SettingsPage.swift, bound to @AppStorage("customApiEndpoint"). Placeholder text is https://your-proxy:8766. Help text reads: "Route API calls through a custom endpoint (e.g. corporate proxy, GitHub Copilot bridge). Leave empty to use the default Anthropic API."
Reading an April 11-12 roundup vs reading a repo
Two ways to answer which AI model shipped this weekend. Only one of them tells a Mac user what to do Monday morning.
| Feature | Standard April 11-12 roundup | This guide (Fazm source) |
|---|---|---|
| Claude Mythos 5 | Listed with parameter count. No picker guidance. | Mapped to Smart label once ACP SDK exposes it. Also subject to Opus-style rate limit concern. |
| Gemma 4 (Apache 2.0) | Listed as open-source. No routing story. | Runnable via Ollama + Anthropic-shaped bridge, routed through ANTHROPIC_BASE_URL. |
| Gemini 3.1 Ultra (2M ctx) | Headline spec mentioned, nothing else. | Not Anthropic-shaped by default, needs a gateway adapter. Noted. |
| TurboQuant paper | Abstract quoted, implications not translated. | Explained as economics shift for long-ctx serving, no user action. |
| Picking a model on Monday | Benchmarks. No decision framework. | Three labels: Scary, Fast, Smart. Smart is auto-downgraded to Fast for rate-limit reasons. |
| Using an open-source release same day | Not addressed. | Paste endpoint URL into Custom API Endpoint. Restart chat. |
| Verifiability | Press releases and benchmark screenshots. | Every claim maps to a file path and line range in the Fazm repo. |
Roundups are useful for knowing what exists. The repo is how you know what to do about it.
The weekend of April 11-12, 2026, in order
Mostly big-lab events, one Mac-app event in the middle. Time pegged to the most specific public signal available (paper upload, blog post, release tag).
Claude Mythos 5 announcement
Anthropic positions Mythos 5 as the first widely recognized 10-trillion-parameter model, with emphasis on long-horizon autonomy.
Gemma 4 release + Gemma 3 license refresh
Google posts Gemma 4 under Apache 2.0 (9B and 27B), and on the same day drops the user-count cap on the Gemma 3 commercial license.
Fazm v2.2.0 ships on Mac
Custom API endpoint setting for proxies and corporate gateways. Global pop-out chat shortcut. Paywall copy tightened.
Gemini 3.1 Ultra surfaces with 2M context
Google DeepMind details a single training objective over text, audio, image, video. 2-million token native multimodal context.
TurboQuant paper at ICLR 2026
Google Research publishes the PolarQuant + Quantized Johnson-Lindenstrauss method for KV-cache compression.
Fazm v2.2.1 ships on Mac
Single-line release: fixes duplicate AI response appearing in pop-out and floating bar when sending follow-up messages.
The number to remember
0 labels, 0 Swift lines, 0 env var.
Three hardcoded model labels in ShortcutSettings.swift. Two lines in ACPBridge.swift that set ANTHROPIC_BASE_URL. One env var that makes every April 11-12 open-source model routable from the same Mac chat. That is the entire surface this guide is about.
Route your Mac chat to the model that dropped this weekend
Install Fazm, open Settings, find the Custom API Endpoint field, paste the URL of any Anthropic-shaped inference server (Claude, Ollama bridge, vLLM bridge, corporate gateway). The floating bar starts sending to it on the next turn. The same three labels still frame the default choice.
Try Fazm free →Frequently asked questions
Frequently asked questions
What actually shipped in AI on April 11 and April 12, 2026?
Four things the big outlets agreed on. Anthropic's Claude Mythos 5, the first widely recognized ten-trillion-parameter model. Google's Gemma 4 open-weight series under an updated Apache 2.0 license that removed the old user-count cap. Google DeepMind's Gemini 3.1 Ultra with a native 2-million token multimodal context. And a TurboQuant paper at ICLR 2026 from Google Research that cuts KV cache overhead using PolarQuant rotation plus a Quantized Johnson-Lindenstrauss compression step. Seven open-source model drops in the first twelve days of April, according to LLM-Stats, landed alongside these headliners.
What did Fazm ship on April 11 and April 12, 2026 specifically?
Fazm v2.2.0 shipped April 11, 2026. The headline change in the CHANGELOG.json file in the Fazm repo reads: Added custom API endpoint setting for proxies and corporate gateways. That single line is why April 11-12 matters for Fazm users, it makes every open-source model that dropped that weekend routable from the same Mac chat with no app update. Fazm v2.2.1 followed on April 12, a bug fix for duplicate AI responses appearing in the pop-out chat and the floating bar when sending follow-up messages. Both versions live in /Users/matthewdi/fazm/CHANGELOG.json with the dates 2026-04-11 and 2026-04-12.
Where in the Fazm source are the three model options actually declared?
File Desktop/Sources/FloatingControlBar/ShortcutSettings.swift, lines 150 to 155. The struct is called ModelOption. The three defaults are hardcoded in this order: Scary (Haiku) = claude-haiku-4-5-20251001, Fast (Sonnet) = claude-sonnet-4-6, Smart (Opus) = claude-opus-4-6. The labels are the labels users see in the floating bar. They are not A/B tests, not experiments, not feature flags. They are three string literals in a Swift array that anyone can read. This is the entire model taxonomy a consumer Mac app committed to the weekend of April 11-12, 2026.
Why does the Fazm source migrate users away from Opus?
Because Opus, the label Fazm calls Smart, burns through rate limits too fast for the loops the agent runs. The exact comment is on line 287 of ShortcutSettings.swift: "One-time migration: switch existing Opus users to Sonnet (Opus burns through rate limits too fast)". The migration runs in the initializer at lines 290 to 293. If a user had claude-opus-4-6 as their stored selection and the migration flag was not yet set, Fazm rewrites their selection to claude-sonnet-4-6, sets shortcut_didMigrateFromOpus to true, and moves on. The Smart option still appears in the picker, it just no longer defaults on for users who chose it once and forgot.
How does Fazm route to Gemma 4 or any other open-source model that dropped April 11-12?
Through the Custom API Endpoint setting added in v2.2.0. The UI lives in Desktop/Sources/MainWindow/Pages/SettingsPage.swift around line 933 as a TextField bound to @AppStorage("customApiEndpoint"). The plumbing is in Desktop/Sources/Chat/ACPBridge.swift, lines 379 and 380: if the user filled in a custom endpoint, the ACP subprocess gets env.ANTHROPIC_BASE_URL set to that URL. Any inference server that speaks the Anthropic API shape (Ollama with an Anthropic-compatible proxy, a vLLM bridge, a corporate GitHub Copilot gateway, anything) takes over for Claude. Gemma 4 served behind such a bridge becomes the model Fazm talks to without a Fazm release.
Why are the labels Scary, Fast, and Smart, rather than the model names?
Because most Mac users do not know what Haiku, Sonnet, or Opus are. Fazm also ships a mapping at ShortcutSettings.swift lines 157 to 162, a private shortLabelMap array that maps the substring haiku to Scary, sonnet to Fast, opus to Smart. When ACP reports a new model list dynamically at runtime (updateModels, line 168), any model whose ID contains one of those three substrings gets the matching user-friendly label. Models that do not match fall back to the name returned by the API, stripped of the Claude prefix. The short labels are the entire surface a user sees, the full model IDs live only in logs and UserDefaults.
Does this let me run Gemma 4 locally and still get Mac app automation?
Yes, with a small caveat. Gemma 4 does not speak the Anthropic Messages API natively, so you need a bridge (a local proxy that translates Anthropic-shaped requests into the format your inference server expects). Once that bridge is running (for example on localhost:11434), you paste its URL into the Fazm Custom API Endpoint field, restart the chat session, and Fazm sends the same accessibility-tree-powered tool calls to your local Gemma 4 instance. Latency depends on your hardware. Tool-use reliability on smaller open-source models is generally worse than on Sonnet, that is what the April 11-12 papers, including the TurboQuant work, are trying to close.
Is any of this tied to screenshot-based computer-use agents?
No. Fazm does not send screenshots as primary context when it can avoid it. The Mac app reads the accessibility tree of whatever app is in the foreground through AXUIElementCreateApplication and AXUIElementCopyAttributeValue (the exact two calls are in Desktop/Sources/AppState.swift around line 439), passes the resulting structured text to whichever model is selected, and gets back tool calls that name AX nodes by role and label rather than by pixel coordinate. This matters for April 11-12 model evaluation because a deterministic text surface means you can compare Claude Mythos 5, Sonnet, Haiku, and a Gemma 4 served through your custom endpoint on the exact same input. No screenshot variance. Same tree, different model, different behavior.
What is TurboQuant and does it matter for desktop AI yet?
TurboQuant is a Google Research paper presented at ICLR 2026 on April 11 that reduces KV cache memory using PolarQuant (a vector-rotation pre-step) then a Quantized Johnson-Lindenstrauss projection. For server-side inference the implications are large, longer contexts at lower memory cost means models like Gemini 3.1 Ultra can serve the 2-million-token context more cheaply. For desktop AI today, the concrete effect is indirect: cheaper long-context serving is what makes it practical for a consumer app to stream an entire accessibility tree plus session history on every turn without hitting a token budget. The paper does not ship code a Fazm user runs, it shifts the economics the hosted endpoints run on.
Can I verify the anchor facts in this guide myself?
Yes. The CHANGELOG lives at /Users/matthewdi/fazm/CHANGELOG.json with 2.2.0 dated 2026-04-11 and 2.2.1 dated 2026-04-12. The three model defaults are at lines 152 to 154 of Desktop/Sources/FloatingControlBar/ShortcutSettings.swift. The Opus-to-Sonnet migration comment and code are at lines 287 to 296 in the same file. The ANTHROPIC_BASE_URL env injection is at lines 379 and 380 of Desktop/Sources/Chat/ACPBridge.swift. The Custom API Endpoint TextField that writes the UserDefault is at line 933 of Desktop/Sources/MainWindow/Pages/SettingsPage.swift. Every claim in this guide maps to a specific line range someone reviewing the repo can open.
Comments
Public and anonymous. No signup.
Loading…