Which new AI models from April 2026 (week 14) should developers know about?

In the week of 30 March to 5 April 2026, Google released the open-weight Gemma 4 family (2 April) and Alibaba's Qwen team launched Qwen3.6-Plus, a hosted coding model with a context window of a million tokens (1 April). Smaller models target specific jobs: IBM's Granite 4.0 3B Vision reads documents, H Company's Holo3 operates desktop software, Microsoft's MAI-Transcribe-1 transcribes 25 languages and Netflix's VOID removes objects from video. Five of the releases in this digest have open weights under the Apache 2.0 license.

What is new in Gemma 4?

Google DeepMind released Gemma 4 on 2 April 2026 in four sizes: E2B and E4B for phones and other devices (2.3 and 4.5 billion effective parameters), a 31B dense model and a 26B mixture-of-experts model that activates about 4 billion parameters per token. All four are under the Apache 2.0 license. They accept text, images and video, and only E2B and E4B also take audio. The small models have a context window of 128,000 tokens, the large ones 256,000.

In the Hugging Face launch post, the 31B model has an estimated LMArena text score of 1452, 85.2% on MMLU Pro and 80.0% on LiveCodeBench v6. The mixture-of-experts model reaches an estimated 1441. According to Google, the unquantized 31B and 26B models fit on a single 80 GB NVIDIA H100, and quantized versions run on consumer GPUs. The architecture alternates sliding-window and global attention layers, adds per-layer embeddings, and uses a shared KV cache in which the last layers reuse keys and values from earlier layers.

Which models read documents and screens?

Granite 4.0 3B Vision (IBM, 31 March 2026) is a compact vision-language model for business documents: tables, charts and key-value pairs in forms. It ships as a LoRA adapter, a small set of extra weights (see PEFT), on top of the Granite 4.0 Micro language model, so the same deployment can also serve text-only requests. With DeepStack Injection, abstract visual features go into early layers of the model and high-resolution detail into later layers. In IBM's tests it scores 86.4% on Chart2Summary, 92.1 on PubTablesV2 (cropped tables) and 85.5% exact match on VAREX, a set of 1,777 US government forms. It is under the Apache 2.0 license.

Falcon Perception from TII (1 April 2026) is a model of 0.6 billion parameters that finds and outlines objects in an image from a text prompt. It processes image patches and text in one sequence, an approach called early fusion. On the SA-Co benchmark it scores 68.0 Macro-F1, against 62.3 for SAM 3, but it is weaker at judging whether an object is present at all. Its sibling Falcon OCR (0.3 billion parameters) scores 80.3 on olmOCR and 88.6 on OmniDocBench. Both are under Apache 2.0.

Holo3 from H Company (1 April 2026) operates desktop software through the screen. H Company reports 78.85% on OSWorld-Verified for its largest version (122 billion parameters, 10 billion active) and calls that a new best score. For comparison, OpenAI reported 75.0% for GPT-5.4 in March. H Company trained Holo3 in synthetic company environments that coding agents build from scenario descriptions. Only the smaller Holo3-35B-A3B has open weights under Apache 2.0; the large version is available through H Company's API.

What did Alibaba's Qwen team release?

Qwen3.6-Plus (1 April 2026) is a hosted model on Alibaba Cloud Model Studio with a context window of 1 million tokens by default and a focus on agentic coding. In Qwen's own table it scores 56.6 on SWE-bench Pro, against 57.1 for Claude Opus 4.5, and 61.6 on Terminal-Bench 2.0, against 59.3 for Opus 4.5. Qwen ran all models on its own corrected version of SWE-bench Pro. OpenRouter offered a free variant from 2 April, according to an archived copy of the page.

Qwen3.5-Omni (29 March 2026) understands text, images, audio and video in one model. It comes in three sizes (Plus, Flash and Light) and recognizes speech in 113 languages and dialects. Qwen says the Plus version beats Gemini 3.1 Pro on audio understanding and matches it on combined audio and video. It is available through Qwen's API.

What else stood out?

• MAI-Transcribe-1 (Microsoft, 2 April 2026) turns speech into text in 25 languages. On the FLEURS benchmark Microsoft measures an average word error rate of 3.9%, against 4.2% for GPT-Transcribe and 7.6% for Whisper-large-v3. It launched in public preview on Microsoft Foundry at $0.36 per hour of audio.
• VOID (Netflix, paper published 2 April 2026) removes objects from video together with the effects they cause, such as objects that fall when a person is removed from the scene. It builds on the CogVideoX video model (5 billion parameters), needs a GPU with at least 40 GB of memory and is under Apache 2.0.
• TTT-E2E, from the Astera Institute, NVIDIA, Stanford, UC Berkeley and UC San Diego, lets a 3-billion-parameter model update part of its weights while it reads, so the time per generated token stays constant as the context grows to 128,000 tokens. The paper dates from December 2025; The Batch covered it on 3 April 2026. The catch: in a needle-in-a-haystack test at 128,000 tokens it scores 6%, against 99% for a standard transformer.
• Lyria 3 (Google) generates music with vocals. It has been in the Gemini app since 18 February 2026 and in public preview for developers through the Gemini API since 25 March. Lyria 3 Pro makes tracks of up to about three minutes and Lyria 3 Clip makes 30-second clips. Every track carries a SynthID watermark, and according to The Batch, Google licensed the training data.