Ali Asif

The AI log, frontier · open weights · agents and coding

Covers 27 Apr 2026 to 3 May 2026

Grok 4.3, Mistral Medium 3.5 weights, and the Qwen3.6 family

xAI ships Grok 4.3 with a 1M window, Mistral releases a 128B open model, and Alibaba and NVIDIA add open mid-size models.

By , published

What shipped

  1. xAI

    Grok 4.3

    A 1M-token context taking text and images, five reasoning levels, and $1.25 in and $2.50 out per million tokens under 200K-token prompts.

  2. Mistral AI

    Mistral Medium 3.5

    A 128B dense open-weights model with a 256K context that folds instruction following, reasoning and coding into one set of weights.

  3. Alibaba Qwen

    Qwen3.6-35B-A3B

    The first open-weight Qwen3.6 model, a 35B mixture-of-experts with only 3B parameters active, pitched at agentic coding.

  4. NVIDIA

    Nemotron 3 Nano Omni

    An open 30B-A3B model that takes video, audio, image and text and acts as the perception layer for agent systems.

  5. Mistral AI

    Mistral Workflows

    A public-preview orchestration layer built on Temporal, with durable execution, step tracking and a human approval pause.

  6. arXiv

    GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

    The technical write-up for Z.ai's multimodal agent model.

  7. arXiv

    Recursive Multi-Agent Systems

    The week's top Hugging Face paper, on multi-agent systems with a recursive structure.

The week closed with a crowd of mid-size open models, and I think they are the quiet story.

Grok 4.3

xAI's Grok 4.3 closed out April. The model page lists a one-million-token context, text and image input, and reasoning levels from none to xhigh with low as the default. Prices are $1.25 input and $2.50 output per million tokens for prompts under 200K, doubling at or above that size, with a 20% batch discount. xAI says it leads on agentic tool calling and instruction following, which is its own claim. I would call it a price move as much as a model move.

Mistral Medium 3.5

Mistral released Medium 3.5 around April 26 to 28 as open weights under a modified MIT license. It is a 128B dense model with a 256K context that takes text and images, and the reasoning effort can be set per request, so a quick chat reply and a long agent run can use the same weights. Mistral reports 77.6% on SWE-Bench Verified, ahead of its own Devstral 2. It is also priced at $1.50 and $7.50 per million tokens on Mistral's API. The same week Mistral put Workflows into public preview, an orchestration layer where developers write Python, every step is tracked in Studio, and a single call pauses a run for human approval.

Qwen and NVIDIA go small and open

Alibaba's Qwen3.6-35B-A3B is the first open-weight model in the 3.6 line: 35B total, 3B active, aimed at agentic coding. Alibaba says the weights are on Hugging Face and ModelScope and that the same model runs on its API as qwen3.6-flash, so you can test it hosted before you download anything. NVIDIA's Nemotron 3 Nano Omni arrived April 28 as an open 30B-A3B hybrid with a 256K context. It reads video, audio, images and text, and NVIDIA positions it as the perception sub-agent inside larger agent systems, claiming up to 9.2 times more effective capacity on video reasoning than comparable open omni models. That figure is NVIDIA's.

Papers

Z.ai published the technical write-up for GLM-5V-Turbo, and Recursive Multi-Agent Systems led the weekly Hugging Face list.

What I'd watch next: whether Medium 3.5's four-GPU claim holds up when someone loads a full 256K context on real hardware.

Sources

Topics: arXiv, Qwen, Alibaba Qwen, NVIDIA, Mistral AI, xAI, Mistral, Grok