Pins matching tag:Multimodal

Alibaba open-sources the flagship Qwen3-VL-235B-A22B in Instruct and Thinking versions, a 235B vision-language mixture-of-experts with a 256K-token native context that Alibaba says matches or exceeds Gemini 2.5 Pro on major visual-perception benchmarks.

Mistral AI releases Mistral 3: Mistral Large 3, a 675B-parameter sparse mixture-of-experts model with 41B active parameters under Apache 2.0, and three dense Ministral 3 models (3B, 8B and 14B).

Z.ai releases GLM-4.6V, a 106B vision-language mixture-of-experts model, and the 9B GLM-4.6V-Flash under the MIT license, the first GLM-V models with native function calling, trained for a 128K-token context and priced at $0.30 per million input tokens.

Moonshot AI releases Kimi K2.5, an open-weight native multimodal model continually pre-trained on about 15 trillion mixed visual and text tokens, with an Agent Swarm mode that coordinates up to 100 sub-agents across 1,500 tool calls.

Alibaba releases Qwen3.5-397B-A17B, an open-weight native multimodal model that activates 17B of 397B parameters, plus the hosted Qwen3.5-Plus with a 1M-token context, on the eve of Lunar New Year.

Alibaba's Qwen team releases three open-weight Qwen3.5 mid-size models under Apache 2.0, a 122B mixture-of-experts with 10B active, a 35B with 3B active and a 27B dense model, which it says beat OpenAI's GPT-5 mini on most benchmarks.

Z.ai (Zhipu AI) releases GLM-5V-Turbo, its first multimodal coding foundation model, which turns design mockups, screenshots and video into runnable front-end code and scores 94.8 on Design2Code against Claude Opus 4.6's 77.3.

Alibaba releases Qwen3.6-Plus, a hosted multimodal model aimed at agentic coding with a 1M-token context window, priced at CNY 2 per 1M input tokens, scoring 78.8 on SWE-bench Verified.

Meta releases Muse Spark, the first model from Meta Superintelligence Labs and its first frontier model without open weights; it powers the Meta AI app and scored 52 on the Artificial Analysis Intelligence Index at launch, behind only Gemini 3.1 Pro, GPT-5.4 and Claude Opus 4.6.

Moonshot AI open-sources Kimi K2.6, a 1-trillion-parameter multimodal mixture-of-experts model with 32 billion active parameters, a 256K token context window and agent swarms of up to 300 sub-agents running 4,000 coordinated steps.

Xiaomi releases MiMo-V2.5-Pro, a 1.02T-parameter agentic model with 42B active parameters and a 1M-token context, and MiMo-V2.5, a 310B omnimodal sibling, open-sourcing both under the MIT licence.

Mistral AI releases Mistral Medium 3.5, its first merged flagship: a 128B dense open-weight model with a 256K context window that scores 77.6% on SWE-Bench Verified and costs $1.50 per million input tokens and $7.50 per million output tokens.

MiniMax releases MiniMax-M3, a natively multimodal model with a 1M-token context built on its new sparse attention, scoring 59.0% on SWE-Bench Pro and priced at $0.30 per 1 million input tokens after a launch discount.

Alibaba releases Qwen3.7-Plus, a proprietary multimodal agent model that reads images and video, operates screens and scores 79.0 on ScreenSpot Pro, priced at about $0.40 per 1M input tokens.

Meta releases Muse Spark 1.1, a multimodal reasoning model for agentic tasks with a 1M-token context window, together with a public preview of the Meta Model API at $1.25 per 1 million input tokens and $4.25 per 1 million output tokens; it scores 51 on the Artificial Analysis Intelligence Index.

Thinking Machines Lab releases Inkling, its first model and an open-weights 975B-parameter mixture-of-experts system (41B active) with a 1M-token context that reasons over text, images and audio, built as a base for fine-tuning.

Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open-weight multimodal model with a 1 million token context window, priced at $3.00 per million input tokens and $15.00 per million output tokens, with weights published on 27 July.

Thinking Machines Lab releases Inkling-Small, an open-weights 276B-parameter model with 12B active parameters that matches or beats the 975B Inkling on reasoning and agentic coding benchmarks, including 80.2% on SWE-bench Verified.

Alibaba launches Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts flagship with a 1M-token context window, fifth in Text Arena and second in Vision Arena, and the first Max-tier Qwen whose weights it then released.

Alibaba's Qwen team open-sources Qwen3.8-27B, a dense multimodal 27B model under Apache 2.0 that it says outperforms the proprietary Qwen3.7-Plus on coding and office workflows, with a 262K native context.

Z.ai releases GLM-5.3-Flash, a 320B-parameter (18B active) natively multimodal open-weight model on a newly trained base that it says beats GLM-5.2 at one-tenth of the price, listed at $0.15 per million input tokens.

StepFun launches Step 5 Preview, a 600B-parameter sparse mixture-of-experts flagship with 27B active parameters and a 1M-token context, priced at $1.00 per 1 million input tokens and scoring 44 on the Artificial Analysis Intelligence Index.

Xiaomi releases and open-sources MiMo-V2.6-Pro and MiMo-V2.6-Flash, natively omnimodal models with 1M-token context; Pro scores 46.32 on the Artificial Analysis Intelligence Index, the highest of any open-source model.

Mistral AI releases Mistral Large 4 in public preview, a natively multimodal mixture-of-experts model of about 1 trillion parameters trained on its own cluster, with open weights promised by the end of October 2026.