AI news, without the noise.

Xiaomi released MiMo V2.6 on September 22, 2026, a family of open-weight, natively omnimodal language models that handle text, images, video, and audio with a 1 million-token context window. Built around the slogan "Scaling Reinforcement Learning Toward Self-Improvement," the lineup includes Pro, Flash, Pro UltraSpeed, and a 9B distill, all under an MIT license and available on Hugging Face, ModelScope, Xiaomi's API, and OpenRouter.

Pro is a 1.02-trillion-parameter sparse MoE model with 42B active parameters, Flash is a 309B/15B-active variant, and UltraSpeed targets low latency. Xiaomi reports Pro leading peers on AutomationBench v1.0.6 (53.1), Terminal Bench 2.1 (89.9), and CyberGym (94.0), while trailing on harder coding and exploit benchmarks like Terminal Bench 4.0 (34.9). Artificial Analysis places Pro at an Intelligence Index of 46 at roughly $0.13 per task, with OpenRouter pricing of $0.14/$0.28 per million tokens for Flash and $0.435/$0.87 for Pro.

Alibaba Group unveiled new artificial intelligence chip technologies and ambitious model plans at its annual conference in Hangzhou, including what CEO Eddie Wu described as China's most powerful AI chip. The company also outlined plans to train a far larger AI model and expand its cloud computing infrastructure amid surging demand.

Wu said the new Zhenwu V900 chip delivers three times the performance of its predecessor Zhenwu M890, and Alibaba plans to train a model at 5 to 10 trillion parameters, up from 2.4 trillion in its Qwen3.8-Max. Alibaba expects to exceed 20 gigawatts of computing capacity by 2032, though global supply chain shortages are slowing scaling, while Chinese-designed chips are gaining ground as US export curbs limit Nvidia sales to China.

Unite.AI reported on September 21, 2026 that SpaceXAI released Grok 4.7, priced at $2 per million input tokens and $6 per million output tokens, matching Grok 4.6. The article describes Grok 4.7 as twice as fast at half the price of comparable models, with a larger base than Grok 4.6, a longer reinforcement learning run, and native training on the Grok Bot harness introduced August 11, 2026. Improvements target longer context handling, self-verification, and document/presentation generation.

Benchmark figures from SpaceXAI's published table place Grok 4.7 at 46.3% on CursorBench 4.0 versus 40.4% for Grok 4.6, 41.7% for GPT-5.6 Sol, and 51.8% for Fable 5.1 at xhigh effort. The article also notes that SpaceXAI extended Grok 4.6 to GitHub Copilot, Amazon Bedrock, the Gemini Enterprise Agent Platform, and Microsoft Foundry in August 2026, framing Grok 4.7 as a successor for long-running coding and agentic workloads.

Qwen has open-sourced Qwen-Image-2.1, a compact image model unifying text-to-image generation and image editing in a single system. The visual generation component uses 7B parameters across 32 Single-Stream DiT layers, delivering quality comparable to larger open-source and closed-source models per Qwen-Image-Bench results.

Key improvements include native transparent image generation and editing, a mixed-granularity attention architecture with token-level causal masking for text and chunk-level masking for images plus KV cache reuse for efficiency, and support for up to 10 reference images with local edit specification, improved typography, portrait lighting, and fine detail rendering.

Qwen3.8-Omni-Flash targets workflows including video editing, music video creation, film commentary, audio-visual summarization, and real-time conversations, with a 1M-token context window. Qwen reports an average score improvement exceeding 25% across 29 evaluations versus Qwen3.5-Omni-Plus, and claims audio input API costs fell over 98% while audio-visual input costs dropped over 93%.

For long videos, the model's agent selectively gathers relevant evidence rather than processing every frame, boosting OmniVideoBench accuracy from 63.4 to 67.8 while cutting token use ~45.7%. It also supports up to one hour of audio-visual meeting input, identifying speakers, transcribing, generating minutes, extracting action items, and triggering downstream tool actions such as email or coding tasks based on meeting context.

PrismML released Ternary Bonsai 2 27B, a ternary-weight derivative of Qwen3.8 27B that compresses the 53.80 GB FP16 model to just 5.93 GB while retaining 98.2% of the parent's performance across 20 benchmarks. The 27.36B-parameter architecture keeps Qwen3.8 27B's hybrid attention backbone (roughly 75% linear, 25% full) and supports text plus image input at a 262K-token context length.

Released under Apache 2.0, the quantized model targets local deployment on a 16 GB laptop or single RTX 5090 GPU, requiring PrismML's llama.cpp fork or MLX runtime. PrismML demonstrates it driving Cline coding agents and computer-use workloads locally, positioning the ternary variant as a practical quantization step two months after the original Bonsai 27B, which retained about 95% of Qwen3.8 27B.

TechFlow's September 18, 2026 newsletter reports Zhipu's official launch of GLM-5.3-FlashX, confirming the model was previously previewed to global developers under the codename "Ox Alpha." Zhipu stated the 200 tokens/s speed was achieved through increased infrastructure investment and inference optimization on a computing cluster backed by 100,000 domestically produced chips, with the FlashX API going live under the identifier "GLM-5.3-FlashX."

The launch framing positions FlashX as combining intelligence, price, and speed simultaneously for the first time within the GLM-5 series. The article notes rapid prior adoption of the underlying GLM-5.3-Flash among global developers drove Zhipu to scale inference capacity, making FlashX a direct response to enterprise demand for higher-throughput, lower-latency access to the same model weights.

Mozilla added Mistral Small 4 as a model option in its Firefox Smart Window AI assistant, launching the feature in France with official French-language support. Mozilla cited the model's multilingual performance as a key selection factor, and Smart Window partners including Mistral agree to zero data retention with conversations not saved on Mozilla servers by default.

This partnership gives Mistral a consumer route beyond its predominantly enterprise and government sales, while Mozilla frames the deal as a stand against AI gatekeepers and a commitment to open, user-controlled AI in browsing. Smart Window remains optional in the browser, and UK and Germany launches are expected later in 2026 following the France rollout.

DeepSeek unveiled DeepSeek-V4.1-Flash, a Mixture-of-Experts model with 552 billion backbone parameters, 8 billion active on input and 16 billion active on output. The new model sharply reduces both active parameters and KV-cache size relative to V4-Pro while posting agentic and coding performance ahead of that flagship on several benchmarks.

DeepSeek announced it will consolidate its existing V4-Pro API service into V4.1-Flash, making the new model a practical successor for production traffic. At maximum reasoning effort, V4.1-Flash outperformed GPT-5.6 Sol and Claude Opus 5 on Terminal-Bench 2.1, with additional claimed advantages on DeepSWE v1.1 and AutomationBench. Outside agentic and coding tasks, however, its relative edge is less pronounced.

The Artificial Analysis article positions Ling-3.0-flash-VL as a comparator while covering the finance-focused Ling-3.0-flash-Fin release. Ling-3.0-flash-VL scored 25 on the Artificial Analysis Intelligence Index and shares a 256K token context window with the Fin sibling built on the same MoE base.

On the Finance & Accounting Index, Ling-3.0-flash-Fin matched Ling-3.0-flash-VL at 24, with the Fin variant achieving higher business knowledge accuracy (17% vs. 11%) but also higher hallucination (33% vs. 19%). Ling-3.0-flash-VL scored 1225 Elo on GDPval-AA v2 and 986 on AA-Briefcase for professional agent tasks, and 16% on AutomationBench-AA.

Ant Group's inclusionAI lab released Ling-3.0-flash-VL on 10 September 2026 as an open-weights vision-language model under the MIT licence. It is a mixture-of-experts design with 124 billion total parameters, activating 5.5 billion per token, and ships alongside fp8, fp4, and int4 quantizations. The model accepts text, images, and video inputs and holds a 262,144-token context window, making it the multimodal counterpart to the earlier text-only Ling-3.0-flash.

The architecture combines a ViT encoder with a small two-layer projector and VideoRoPE positional encoding, enabling timestamp-aware queries over long video. Its 42-layer backbone alternates Kimi Delta Attention and Gated MLA at a five-to-one ratio, keeping long-context memory costs manageable. Thinking mode is on by default but can be disabled, and inclusionAI positions the release as strengthening language reasoning while adding visual perception and visual-agent capabilities over its text predecessor.