←Search results

Search “Reasoning”

By relevance · Total 552 results Clear search

Sunday

  1. Story1 reports · 1 facts

    Qwen Releases Omni-Modal Model Qwen3.8-Omni-Flash

    On September 18, 2026, Qwen released the omni-modal model Qwen3.8-Omni-Flash. According to official information, this is its first omni-modal model built around agent capabilities, natively supporting audio-visual understanding, reasoning, and tool calling. The company says its audio-visual capabilities are close to Gemini 3.8 Flash, with agent performance improving by an average of 19.5 points on WildClawBench-MM and UniClawBench; it supports a 1M token context and uses 51.8% fewer tokens than static understanding on OmniVideoBench. Video input costs are about 89% lower than Qwen3.5-Omni-Plus. Alongside the release, Qwen-MM-Plugins is being open-sourced, and Qwen-Live Harness is coming soon. There are currently no subsequent independent development facts.

    On September 18, 2026, Qwen released the omni-modal model Qwen3.8-Omni-Flash. According to official information, this is its first omni-modal model built around agent capabilities, natively supporting audio-visual understanding, reasoning, and tool calling. The company says its audio-visual capabilities are close to Gemini 3.8 Flash, with agent performance improving by an average of 19.5 points on WildClawBench-MM and UniClawBench; it supports a 1M token context and uses 51.8% fewer tokens than

    Matched fieldsSummary

  2. Story1 reports · 1 facts

    Bailing releases and open-sources native multimodal model Ling-3.0-flash-VL

    On September 9, 2026, Ant Bailing released and open-sourced Ling-3.0-flash-VL, the first native multimodal large model in the Ling series. Built as an extension of the Ling-3.0-flash MoE architecture, the model has 124B total parameters and 5.5B activated per inference, natively supports image, text, and video inputs, and has a 256K-token context window. According to the company, native multimodal joint training improved the model's score on the Artificial Analysis Intelligence Index v4.1.1 by 4 points over the text-only version, and it introduces a visual feedback loop of observe, act, verify, and correct, used for front-end replication, GUI agents, and medical report interpretation. The model is currently available for free trial on Ling Studio, and the BF16 and FP8 versions have been open-sourced on Hugging Face and ModelScope.

    On September 9, 2026, Ant Bailing released and open-sourced Ling-3.0-flash-VL, the first native multimodal large model in the Ling series. Built as an extension of the Ling-3.0-flash MoE architecture, the model has 124B total parameters and 5.5B activated per inference, natively supports image, text, and video inputs, and has a 256K-token context window. According to the company, native multimodal joint training improved the model's score on the Artificial Analysis Intelligence Index v4.1.1 by 4

    Matched fieldsSummary

  3. Story1 reports · 1 facts

    Ant Ling officially open-sources Ling-3.0-tiny

    On August 11, 2026, Ant Ling officially open-sourced Ling-3.0-tiny. According to the official introduction, this is a natively hybrid reasoning model with 7.9B total parameters, activating only 1.3B parameters during inference, and it is simultaneously offered in three versions: BF16, FP8, and INT4. The company states that it scores 25 on the Artificial Analysis Intelligence Index and 16 on the Agentic Index, with an output speed exceeding 160 tokens/s. In terms of deployment, the model can perform FP8 inference on a single NVIDIA DGX Spark, with weights of approximately 7.85GB, and has also been adapted for local operation on Apple Silicon MacBook and Mac mini. The model is currently in an open-source release state, with no further developments at this time.

    On August 11, 2026, Ant Ling officially open-sourced Ling-3.0-tiny. According to the official introduction, this is a natively hybrid reasoning model with 7.9B total parameters, activating only 1.3B parameters during inference, and it is simultaneously offered in three versions: BF16, FP8, and INT4. The company states that it scores 25 on the Artificial Analysis Intelligence Index and 16 on the Agentic Index, with an output speed exceeding 160 tokens/s. In terms of deployment, the model can perf

    Matched fieldsSummary

  4. Story1 reports · 1 facts

    Ant Ling officially open-sources the Ling-3.0-flash model

    On August 7, 2026, Ant Ling officially open-sourced Ling-3.0-flash, a new-generation native hybrid reasoning model. According to the Ant Ling developer blog, the model uses a MoE architecture with 124B total parameters and 5.1B activated parameters, and is released alongside a base version as well as multiple precision versions including FP8, FP4, and INT4. There are three deployment options: cloud API calls; end-to-end local inference on a single NVIDIA DGX Spark using the MXFP4 or INT4 version; and a high-performance deployment that exceeds an average output rate of 1,100 tokens/s under the specified GPU test configuration. In terms of benchmark data, its output speed in Artificial Analysis is 353 tokens/s, and in the AA Intelligence Index its weighted average invocation cost per task is about $0.04. There are currently no subsequent development facts.

    On August 7, 2026, Ant Ling officially open-sourced Ling-3.0-flash, a new-generation native hybrid reasoning model. According to the Ant Ling developer blog, the model uses a MoE architecture with 124B total parameters and 5.1B activated parameters, and is released alongside a base version as well as multiple precision versions including FP8, FP4, and INT4. There are three deployment options: cloud API calls; end-to-end local inference on a single NVIDIA DGX Spark using the MXFP4 or INT4 version

    Matched fieldsSummary

Sunday

  1. Story2 reports · 1 facts

    StepFun releases the StepAudio 3 Music music generation foundation model

    On September 16, 2026, StepFun released the StepAudio 3 series of audio models, which includes the music generation foundation model StepAudio 3 Music. According to official information, the series comprises 5 audio models in total, covering real-time speech, speech recognition, speech generation, audio generation, and music. The company also disclosed the performance of the other models in the series: Realtime reached 98.9% and 99.7% on Artificial Analysis's Conversational Dynamics and Speech Reasoning respectively, ranking first in both; ASR's word error rate is 1.7%, matching the best score on the leaderboard. The company states that the series can be used to build voice agents capable of handling interruptions, reasoning while speaking, and calling tools, and it is now available through the Voice AI Lab and blog, with the API also open. There are currently no subsequent development facts.

    On September 16, 2026, StepFun released the StepAudio 3 series of audio models, which includes the music generation foundation model StepAudio 3 Music. According to official information, the series comprises 5 audio models in total, covering real-time speech, speech recognition, speech generation, audio generation, and music. The company also disclosed the performance of the other models in the series: Realtime reached 98.9% and 99.7% on Artificial Analysis's Conversational Dynamics and Speech R

    Matched fieldsSummary

  2. Story2 reports · 1 facts

    Google releases Gemini 3.8 Live audio model

    Google released two real-time conversational audio models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, available starting today through the Gemini API, Google Workspace, and the Gemini app. According to Google DeepMind, Gemini 3.8 Live Extended Thinking ranks first on the Artificial Analysis Speech-to-Speech Quality Index with a score of 82.6, achieves a 68.6% task completion rate on τ-Voice agent tasks, and scores 97.7% on Big Bench Audio; Gemini 3.8 Live ranks second on the Speech Agent Arena. Both models support near-real-time visual input, automatic switching between 97 languages during a conversation, and can execute tool and API calls in the background without interrupting the conversation. On September 18, Google demonstrated a workshop assistant built with Gemini 3.8 Live Extended Thinking that can analyze the workspace, reason aloud, and guide the building process. On September 21, Google also demonstrated a programming tutoring Agent built with this model that can view the user's screen, reference the p5.js library through function calling, and explain code logic.

    Google released two real-time conversational audio models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, available starting today through the Gemini API, Google Workspace, and the Gemini app. According to Google DeepMind, Gemini 3.8 Live Extended Thinking ranks first on the Artificial Analysis Speech-to-Speech Quality Index with a score of 82.6, achieves a 68.6% task completion rate on τ-Voice agent tasks, and scores 97.7% on Big Bench Audio; Gemini 3.8 Live ranks second on the Speech A

    Matched fieldsSummary

  3. Story1 reports · 1 facts

    Fireworks evaluates DeepSeek's new model

    On September 14, 2026, Fireworks AI released and evaluated DeepSeek-V4.1-Flash. According to its published results, the model achieves 74.34% pass@1 on DeepSWE, on par with GPT-6-Astra's 74.12%; its cost per task is $0.43, about 1/15 that of Astra. The model is a 552B-parameter MoE, using separate activation budgets of 8B for input and 16B for output, and its KV Cache HBM footprint is reduced to 1/4 that of the previous-generation V4. On HLE it scores 34.52%, below Astra's 50.40%, and Fireworks considers it better suited to agentic coding than to highly difficult reasoning. On September 28, Fireworks AI released its first cache-aware router, FireRouter with Opus, which is now live in the CLI and offered for the first time as a standalone routing model; it can select among Claude Opus 5.5, GLM 5.3 and GLM 5.3 Flash based on task fit, prompt cache cost and switching cost; its internal A/B testing over more than a month showed coding task cost per session dropping from $15.36 to $6.63, a 57% reduction (±19 percentage points), with accuracy at pure O

    On September 14, 2026, Fireworks AI released and evaluated DeepSeek-V4.1-Flash. According to its published results, the model achieves 74.34% pass@1 on DeepSWE, on par with GPT-6-Astra's 74.12%; its cost per task is $0.43, about 1/15 that of Astra. The model is a 552B-parameter MoE, using separate activation budgets of 8B for input and 16B for output, and its KV Cache HBM footprint is reduced to 1/4 that of the previous-generation V4. On HLE it scores 34.52%, below Astra's 50.40%, and Fireworks

    Matched fieldsSummary

  13. Story2 reports · 1 facts

    Fireworks launches Muse Glimmer

    On August 10, 2026, Fireworks AI launched serverless and on-demand deployment services for Meta's Muse Glimmer 30B dense model. The model is released under the Apache 2.0 license, supports a 128K+ token context window and native image understanding, and targets persistent agent scenarios that require multi-turn tool calling. Architecturally, it uses sliding window attention, with a window of 2,048 tokens in most layers and one global attention layer every four layers, combined with 2 KV heads to control KV cache usage. Fireworks provides recommended parameters of temperature 1.0, top_p 0.95, and top_k 64, and supports setting reasoning effort from low to xhigh in the system prompt. According to Meta's report, the model scores 75.5 on MCP Atlas, 74.6 on DeepSearch QA, and 51.2 on SWE-Bench Pro, all higher than the corresponding results of Gemma 4 31B and Qwen 3.6 27B. There are currently no further developments.

    On August 10, 2026, Fireworks AI launched serverless and on-demand deployment services for Meta's Muse Glimmer 30B dense model. The model is released under the Apache 2.0 license, supports a 128K+ token context window and native image understanding, and targets persistent agent scenarios that require multi-turn tool calling. Architecturally, it uses sliding window attention, with a window of 2,048 tokens in most layers and one global attention layer every four layers, combined with 2 KV heads to

    Matched fieldsSummary

Wednesday

Thursday