←Search results

Search “Model releases”

By relevance · Total 25 results Clear search

Wednesday

  1. ModelBest OpenBMB

    MiniCPM5-2B-Jev Decision Model Released: 78.79% on JevBench

    Developer ActbroDev built the System 1 decision model MiniCPM5-2B-Jev based on OpenBMB's MiniCPM5-2B. After inputting a document and a question, it directly returns the probabilities of predefined options rather than generating free text. The project reports that it scores 78.79% on JevBench, ranking first among ≤2B open-weight System 1 / Jev models, with 86.41% accuracy on the decision-v7 benchmark and an ECE of 0.0307. The code, LoRA adapters, and calibration head are open-sourced under Apache 2.0, and it provides a /v1/systemone structured decision API and shared-context KV-cache reuse.

    Developer ActbroDev built the System 1 decision model MiniCPM5-2B-Jev based on OpenBMB's MiniCPM5-2B. After inputting a document and a question, it directly returns the probabilities of predefined options rather than generating free text. The project reports that it scores 78.79% on JevBench, ranking first among ≤2B open-weight System 1 / Jev models, with 86.41% accuracy on the decision-v7 benchmark and an ECE of 0.0307. The code, LoRA adapters, and calibration head are open-sourced under Apache

    Matched fieldsTitle

    #editorial#English text is being prepared.#English text is being prepared.#English text is being prepared.

Wednesday

  1. TechCrunch AI

    Anthropic proposes draft AI safety bill that would block model releases when risk of 50+ deaths exists

    Anthropic has pitched a draft AI safety bill to Congress that would allow a model's release to be blocked when it poses a foreseeable, imminent risk that could cause 50 or more deaths or more than $1 billion in property damage. The risk must stem from the manufacture of bioweapons, the model evading its developer's control, or cyberattacks. The draft requires an independent evaluator registered with the U.S. Department of Commerce to determine the risk; the evaluator must report within 72 hours, after which the Commerce Department would apply to a Washington, D.C., judge for an order to halt the release. Federal agencies would not be allowed to decide on blocking a release themselves.

    Anthropic has pitched a draft AI safety bill to Congress that would allow a model's release to be blocked when it poses a foreseeable, imminent risk that could cause 50 or more deaths or more than $1 billion in property damage. The risk must stem from the manufacture of bioweapons, the model evading its developer's control, or cyberattacks. The draft requires an independent evaluator registered with the U.S. Department of Commerce to determine the risk; the evaluator must report within 72 hours,

    Matched fieldsTitle

    #tech-media#English text is being prepared.#Anthropic#Safety / alignment

Tuesday

  2. FactRelated story

    Qwen3.8-27B RL post-training model released

    On October 20, 2026, an open-source environment model post-trained with reinforcement learning based on Qwen3.8-27B was released. According to Rohan Paul, the model reduced the hallucination rate on AA-Omniscience from 30.4% to 15.1%, and surpassed Muse Glimmer 30B and GPT-6 Luna high on Terminal-Bench, τ3-bench, ClawEval, and BFCL v4; however, on DeepSWE, Luna still leads with 53.1%. Its median output for solved tasks is 23.8K tokens, compared with 54.9K for the base Qwen. The authors also mention that the MLA version replaces 16 layers of full attention with multi-head latent attention, reducing memory usage per token from 64 KiB to 18 KiB; the compressed version can run on a single 24 GB RTX 4090, with a context about 3.5 times that of the base Qwen.

    Matched fieldsTitle

    View story:Qwen3.8-27B RL Post-Training Model Released

Tuesday

  2. David Singleton

    Muse Wonder model released, aimed at personal agents and made available first to subscribers

    The Muse Wonder model has been released, positioned to power personal agents, and has been made available first to subscribers. The model includes three capabilities: Active Memory, Active Focus, and domain expertise. The company says its memory spans day, week, and month scales, and that it achieves SOTA in finance and legal domains and performs strongly in health. David Singleton said he has been using it for some time and believes that under Active Focus and Active Memory the model works harder and calls on memory more precisely.

    The Muse Wonder model has been released, positioned to power personal agents, and has been made available first to subscribers. The model includes three capabilities: Active Memory, Active Focus, and domain expertise. The company says its memory spans day, week, and month scales, and that it achieves SOTA in finance and legal domains and performs strongly in health. David Singleton said he has been using it for some time and believes that under Active Focus and Active Memory the model works hard

    Matched fieldsTitle

    #analysis#Product updates#Agents#OpenAI

Tuesday

Wednesday

Tuesday

Saturday

Tuesday

Wednesday

Thursday

Friday

Wednesday

Thursday

Sunday

  1. Story1 reports · 1 facts

    Qwen3.8-27B RL Post-Training Model Released

    On October 20, 2026, an open-source environment model post-trained with reinforcement learning based on Qwen3.8-27B was released. According to Rohan Paul, the model reduced the hallucination rate on AA-Omniscience from 30.4% to 15.1%, and surpassed Muse Glimmer 30B and GPT-6 Luna high on Terminal-Bench, τ3-bench, ClawEval, and BFCL v4; however, on DeepSWE, Luna still leads with 53.1%. Its median output for solved tasks is 23.8K tokens, compared with 54.9K for the base Qwen. The authors also mention that the MLA version replaces 16 layers of full attention with multi-head latent attention, reducing memory usage per token from 64 KiB to 18 KiB; the compressed version can run on a single 24 GB RTX 4090, with a context about 3.5 times that of the base Qwen.

    On October 20, 2026, an open-source environment model post-trained with reinforcement learning based on Qwen3.8-27B was released. According to Rohan Paul, the model reduced the hallucination rate on AA-Omniscience from 30.4% to 15.1%, and surpassed Muse Glimmer 30B and GPT-6 Luna high on Terminal-Bench, τ3-bench, ClawEval, and BFCL v4; however, on DeepSWE, Luna still leads with 53.1%. Its median output for solved tasks is 23.8K tokens, compared with 54.9K for the base Qwen. The authors also ment

    Matched fieldsTitle

Sunday

Friday