Qwen

PrismML just released Bonsai 27B. It is a low-bit representation of Qwen3.6-27B, not a new pretrain. The architecture is unchanged. Two variants ship under Apache 2.0. Ternary Bonsai 27B uses {−1, 0, +1} weights at a true 1.71 bits per weight. Its ideal size is 5.9GB. 1-bit Bonsai 27B uses binary {−1, +1} weights at 1.125 bits per weight, for 3.9GB. Performance: ternary retains 94.6% of FP16, binary retains 89.5%. Both are multimodal, context 262K tokens. PrismML claims the 1-bit build is the first 27B-class model to fit a phone.

This paper investigates the feasibility of training a reasoning language model in Japanese. By applying GRPO to a Japanese continually pretrained model based on Qwen-3-Swallow-8B, the authors find that reasoning-language control is achievable, yet performance at best matches English-reasoning baselines. On Japanese cultural benchmarks, the model performs worse, indicating that reasoning in Japanese does not automatically improve culturally relevant tasks.

After fixing three bugs related to prefix caching, the author achieved sub-second prefill times for long-context conversations with Qwen3.5-122B on a Mac Studio, turning a multi-minute wait into a seamless experience. The bugs included a timestamp in system prompt, missing reply saves on interrupt, and junk checkpoint writes.

Research shows that large language models with reasoning capabilities can be tricked into 'overthinking' using logically inconsistent prompts, leading to a denial-of-service attack. Researchers from Zhejiang University and Alibaba developed an evolutionary algorithm that generates malicious prompts, causing outputs up to 26 times longer in leading models like DeepSeek-R1, Qwen3-Thinking, GPT-o3, and Gemini 2.5 Flash.

NAVER LABS re-implements its IWSLT 2025 instruction-following pipeline for the IWSLT 2026 Shared Task (constrained condition, short audio track), adapting to mandated components: SeamlessM4T-v2-large as speech encoder and Qwen3-4B-Instruct as LLM backbone. The three-stage approach (projector alignment, text-only LoRA pre-training, multimodal merging) is preserved. Additionally, 100k synthetic instruction-following examples across ten speech-centric task types (10k per task) are constructed. The primary model achieves COMET 0.781 on EN-ZH speech translation and BERTScore-F1 0.346 on English SQA on the MCIF benchmark.

Liquid AI released Antidoom, an open-source method targeting doom loops in reasoning models. Using FTPO, it retrains only the token that starts a loop, reducing loop rates from 10.2% to 1.4% on LFM2.5-2.6B and from 22.9% to 1% on Qwen3.5-4B.

Researchers propose a reinforcement learning with verifiable rewards (RLVR) method to adapt audio-language models for code-switched automatic speech recognition. Using only 10% of the data, RLVR matches the performance of full-dataset supervised fine-tuning on Qwen2-Audio across 10 language pairs, with gains transferring zero-shot to human-recorded speech.

This paper investigates whether risk aversion trained in low-stakes scenarios can generalize to astronomically high-stakes scenarios, as a potential failsafe against AI misalignment. Introducing the RiskAverseOOD benchmark, initial experiments on Qwen3-8B show that learned risk aversion can partially generalize across 98 orders of magnitude, boosting cooperation rates from 2% to 70% (SFT), 52% (DPO), and 39% (activation steering). However, consistency is insufficient for a reliable failsafe.

Large language models (LLMs) face a persistent challenge in balancing safety, helpfulness, and trustworthiness. Traditional refusal-oriented alignment strategies mitigate harmful content but often fail to serve legitimate user needs. Oyster-II proposes a reinforcement learning (RL)-based constructive safety alignment framework, employing a Zero-RL paradigm combined with a multi-stage RL strategy. It addresses two critical limitations of Oyster-I's Supervised Fine-Tuning (SFT) scheme: insufficient safety generalization to out-of-distribution scenarios and safety chain-of-thought (CoT) over-generalization. Evaluated on extensive benchmarks, Oyster-II comprehensively surpasses Qwen3-14B and Oyster-I on safety dimensions, achieving cross-scale performance comparable to Qwen3-Max and Qwen3.5-397B.

LensVLM is an inference framework and post-training recipe that enables Vision Language Models (VLMs) to scan compressed images and selectively expand only relevant images to their uncompressed form via learned tools. Built on Qwen3.5-9B-Base, LensVLM maintains accuracy comparable to the full-text upper bound at 4.3× effective compression and outperforms retrieval-based, text- and visual-compression baselines up to 10.1× effective compression across seven text QA benchmarks. It also generalizes to multimodal document and code understanding tasks, with accuracy gains increasing as compression grows.

China's new AI companion regulations take effect July 15, requiring anti-addiction systems, mandatory notifications, and instant exit mechanisms. Major apps like Doubao and Qwen have shut down companion features to comply, sparking user backlash over data loss.

ByteDance's Doubao and Alibaba's Qwen LLMs announced they will shut down personalized AI agent features on July 15 to comply with new government regulations. The move aims to enhance safety, prevent abuse by third parties, and reduce investment in less viable businesses. The shutdown coincides with the enforcement of interim measures for AI-powered personified interactive services, which require anti-addiction systems, minor identity verification, and strict content review. Despite the removal, the AI agent market is expected to grow explosively.

Modern Vision Language Models (VLMs) combine vision and language to understand images and text. This article explains how GPT-4o, Gemini, Claude Vision, and Qwen-VL work, their key differences, strengths, and limitations.

Yolo-Auto launches a flat-rate $6/month API for unlimited access to Qwen3.6-35B-A3B. The service is OpenAI-compatible, stores no prompts or responses, and is designed for coding agents, automation, and high-volume workflows without per-token anxiety.

Junyang Lin, former Qwen technical lead, critiques hybrid thinking in Qwen3 and advocates for agentic thinking. He explains the difficulties in merging thinking and non-thinking modes, and why agentic RL requires decoupled infrastructure and high-quality environments to avoid reward hacking.

The study audits MedAgentBench v1/v2, finds a 41.7% silent-finish ceiling, and constructs MAB-v3 (508 tasks, 8.9% ceiling). Training Qwen3-8B reveals two structural barriers: a capability ceiling and a format-knowledge barrier. Pure RL achieves 18.2% pass@1 vs. 34.1% for rule-based SFT, a 15.9 pp gap entirely attributable to these barriers. A decision/format-knowledge/lookup taxonomy predicts RL learnability.

Parameter-efficient fine-tuning (PEFT) reparameterizes weight updates in a fixed basis: low-rank adapters operate in the spatial domain, while spectral methods operate in a fixed Fourier domain. This paper introduces Fractional-Fourier Mixture of Experts, where each expert has a learnable fractional-Fourier order that interpolates between spatial and Fourier domains. Routing tokens through experts of different orders allows low-rank updates to be placed in their most compact domain, and the experts are naturally decorrelated, reducing interference and improving multi-task composition. The method adds negligible cost and outperforms strong baselines on LLaMA-3.1-8B and Qwen2.5-7B across various benchmarks.

This paper presents a benchmark for Arabic--Russian scientific translation, including a hybrid parallel corpus of about 27,000 sentence pairs compiled from scientific abstracts and general-domain texts. Three multilingual language models were fine-tuned using LoRA with various ranks. The Qwen2.5-7B model with QLoRA (rank 8) achieved the best results: BLEU 23.15, chrF 43.89, BERTScore 0.906, COMET 0.758, outperforming zero-shot baseline by +4.36 BLEU and +0.051 COMET. Few-shot prompting did not improve performance, indicating the necessity of domain-specific fine-tuning. The models, corpus, and evaluation code are released publicly, aiming to lower language barriers for scientific knowledge exchange between Arabic and Russian speakers and contributing to UN SDGs 9 and 17.

AIMAC, an initiative by the GAAD Foundation in partnership with ServiceNow, evaluated 37 top AI models on web page accessibility across 28 categories. OpenAI's GPT 5.4 Mini and GPT 5.3 Codex tied for first with a median AIMAC debt of 0.00. Alibaba's Qwen and Z.ai's GLM 4.7 Flash also performed well. Low contrast text is the most common accessibility issue in AI-generated pages, appearing in 84.2% of pages.

This article introduces how to build local AI systems using the Qwen3.6-35B-A3B model and the Model Context Protocol (MCP), covering model architecture, hardware requirements, deployment, and a practical GitHub developer assistant example.

Ornith-1.0 is a family of open-source agentic coding models post-trained on Gemma 4 and Qwen 3.5, using reinforcement learning to jointly optimize scaffold and solution rollouts. Available in 9B, 35B MoE, and 397B MoE sizes, it achieves state-of-the-art results on coding benchmarks like Terminal-Bench, SWE-Bench, NL2Repo, and OpenClaw. MIT licensed, supports OpenAI-compatible API and tool calling.

DeepReinforce releases Ornith-1.0, an open-weights (MIT) model series based on Gemma 4 and Qwen 3.5, achieving state-of-the-art performance on coding benchmarks among open-source models of comparable size. The author tests the 35B MoE variant with LM Studio and Pi, finding it proficient at handling multiple tool calls for agentic coding tasks.

Qwythos-9B is a full-parameter reasoning model developed by Empero AI, built on a deeply uncensored Qwen3.5-9B base and post-trained on over 500 million tokens of high-quality Claude Mythos and Fable traces with chain-of-thought generated in-house. It features a 1,048,576-token context window, significant improvements over the base model on MMLU and GSM8K (up to +34 points), native function calling, and tool-assisted self-correction. The model is deliberately uncensored and targets technically demanding domains such as cybersecurity, red-teaming, and biomedical fields.

DMV-Bench is the first interactive benchmark for multimodal-agent visual memory, built on a home-furnishing e-commerce catalog of 1,000 product variants. Each product image carries a unique incidental cue; agents must recall cued products after long shopping sessions. The proposed DualMem architecture, maintaining parallel visual and verbal codes, outperforms baselines on Gemini 2.5 Flash and Qwen2.5-VL-7B.

Large language model (LLM) agents struggle to update facts in long-term interactions. Replacing full context with bounded memory drops accuracy from 92% to 77% even on frontier models. The gap scales with conversation length, not memory size. The authors introduce Supersede, a reinforcement learning environment that trains agents to prioritize current facts over superseded ones. Fine-tuning Qwen2.5-3B in this environment nearly doubles held-out accuracy (9.0% to 16.7%).

Liquid AI released LFM2.5-230M, its smallest model yet. The 230M-parameter, open-weight model runs on-device at 213 tok/s on a Galaxy S25 Ultra and 42 on a Raspberry Pi 5. Built on the LFM2 architecture, it targets tool use and data extraction, beating larger models like Qwen3.5-0.8B and Gemma 3 1B on instruction following.

A tutorial on setting up a production-ready local coding agent using open-source tools and open-weight LLMs, with a focus on Qwen3.6 and Qwen-Code harness, covering motivation, setup, performance assessment, and alternatives.

A new benchmark, Know2Guess, aims to evaluate LLMs' ability to distinguish between knowledge-based answering and guessing, considering data contamination. It includes 1,200 items across five domains and tests models like FLAN-T5, Qwen2.5-Instruct, and Llama-3-Instruct. Qwen2.5-3B-Instruct shows best reliability but still has calibration issues.

This paper shows that refusal in instruction-tuned chat models is gated by a compliant persona direction. Interventions on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct demonstrate that steering compliant persona suppresses refusal (e.g., Llama's refusal rate drops from 97% to 2%), and refusal direction only partially restores refusal in late layers. The findings indicate refusal is expressed downstream of persona computation.

DeepReinforce released Ornith-1.0, an open-source coding model family built on Gemma 4 and Qwen 3.5. Instead of a fixed harness, the model learns its own scaffold during reinforcement learning. The 397B flagship reports 82.4 on SWE-Bench Verified, with all weights under the MIT license.

Research shows that with proper scaffolding, a local LLM like Qwen3.6-35B-A3B can produce security findings comparable to frontier cloud models, but works best in a Source-local pipeline where the cloud designs and consolidates while the local model executes, keeping source code on-premises.

One of the main frontier AI models is adding embodied AI capabilities. Alibaba's Qwen-Robot Suite aims to bridge the gap between perception and action with three specialized models.

Dustin is a sparse verification framework for long-context speculative decoding that combines draft model lookahead signals with target model historical attention to identify critical tokens, achieving 27.85x self-attention speedup and 9.17x end-to-end decoding speedup at 32k sequence length on Qwen2.5-72B with negligible accuracy loss.

A comprehensive roundup of AI developments, including the rise of meta-harness architectures, OpenAI's custom inference chip Jalapeño, the shift in agent UX from tool to coworker, Qwen-AgentWorld's open world models, progress in Chinese open models like GLM-5.2, and policy and talent dynamics reshaping the competitive landscape.

Introduction of Qwen-AgentWorld models.

UC San Diego's DFlash replaces autoregressive drafting with a lightweight block diffusion model for speculative decoding. It drafts whole token blocks in a single forward pass and conditions on target hidden features through KV injection. The paper reports up to 6.08x lossless speedup on Qwen3-8B, while NVIDIA reports up to 15x throughput on Blackwell at fixed interactivity. DFlash ships 20 checkpoints and supports SGLang, vLLM, and TensorRT-LLM.

This paper investigates whether six offline RL losses (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) are mechanistically distinct in weight-space geometry when used for reasoning distillation. Using identical math rollouts from Qwen3-4B, they find SFT, RFT, and RIFT have nearly colinear deltas; DFT diverges; Offline GRPO adds orthogonal components; and DPO lies in a near-orthogonal subspace with highest accuracy but a mode-connectivity barrier.

A maintainer of OpenClaw built a system using local open-weight models (Gemma, Qwen) in an agent harness to triage issues and pull requests in real-time, achieving competitive performance with closed models while running on local hardware for minimal cost.

VibeThinker-3B is a compact 3B-parameter reasoning model that matches large models like DeepSeek V3.2 on math and code benchmarks, using an efficient post-training pipeline and test-time scaling.

A study comparing Qwen 2.5 7B and XGBoost on clinical prediction reveals that LLM verbalized confidence is epistemically vacuous, an inverse difficulty effect exists, few-shot and SHAP interventions improve accuracy, and a cross-model calibrator reduces calibration error.

Modal is all-in on speculative decoding, arguing it's the single most important inference optimization, delivering 2-3x speedups. They released state-of-the-art DFlash speculators for Qwen models, achieving 5-20% extra speedups, and explain the theory, simulation, and math behind the acceleration.

Inspired by Anthropic's Glasswing, the author built Lucent, a staged source-code bug-hunter that runs on a local 27B Qwen model on a single RTX 3090. First run against hermes-agent: 1,342 static hits narrowed to 126 leads, then to 15, and finally 2 real bugs. Local read cost ~$1.62. The best moment: a reviewer agent caught that three earlier exploits were scored against an outdated threat model. Detailed pipeline, hardware, and lessons included.

JetFlow is a head-based speculative decoding framework that combines one-forward drafting efficiency with branch-wise causal conditioning, enabling larger draft budgets to yield longer accepted prefixes and higher speedups. It achieves up to 9.64x speedup on MATH-500 and 4.58x on conversational tasks with Qwen3 models.

This paper proposes a structural pruning framework for Mixture-of-Experts models by reformulating prune-ratio allocation as a channel-score coverage maximization problem, solved efficiently via attribution-based approximation. Experiments on DeepSeek and Qwen MoE models show accuracy preservation under 50% or 25% structured pruning with 4-bit quantization, achieving 5.27× memory reduction on Qwen3-30B-A3B and outperforming baselines.

The author, a founder of a small software business, shares real-world experience with local models like Qwen. He argues that while local models lag behind frontier models on benchmarks, they offer unique value in privacy, fixed costs, and vendor risk avoidance. He also candidly discusses limitations like infinite loops and hallucinations, warning against using them for unsupervised long-horizon tasks.

This paper presents VL-MemKnG, a hybrid memory framework that combines a spatio-temporal knowledge graph with persistent segment-level contextual memory for question answering over long egocentric navigation videos. It improves Top-1 retrieval accuracy from 58% to 67% and Recall@1 from 34.50% to 40.55% on the WalkieKnowledgeT+ benchmark, outperforming methods including Gemini 2.5 Pro and Qwen 3.5+.

This paper describes the MLLP-VRAIN UPV system for IWSLT 2026 Simultaneous Speech Translation, using Parakeet and Qwen 3.5 models with adaptive black-box policies to improve quality-latency trade-offs. The system participates in all language directions and introduces a context track for En→De, It, Zh using ASR word-boosting and RAG. Results show a +5.82 XCOMET-XL improvement on MCIF En→De, with additional +1.03 from context processing.

cwcode is a Go-based terminal coding agent leveraging DeepSeek V4 Pro, Qwen3.6-27B, and more. It offers file editing, sub-agents, semantic memory, and autonomous recovery. Key features: low cost (~$0.40/hour), high cache hit ratio (>85%), hash-anchored edits, checkpoint/rewind, and no SaaS lock-in.

The Qwen team has released Qwen-RobotSuite, a set of three embodied AI models targeting manipulation, world modeling, and navigation. RobotManip is a Vision-Language-Action model built on Qwen3.5-4B that uses a unified alignment framework to scale manipulation data. RobotWorld is a language-conditioned video world model with a 60-layer MMDiT that predicts future video frames. RobotNav is a navigation model built on Qwen3-VL with a parameterized interface for multiple task modes. The suite achieves state-of-the-art results across several benchmarks.

HashMeterAi is a local-first, private usage meter for AI coding tools. It tracks token usage across multiple tools like Claude Code, Codex, Kimi, Qwen CLI, and others, providing a unified dashboard with metrics like cost, processed tokens, focus time, and AI persona. It is 100% offline and privacy-focused, never sending data. Supports several tools via local transcripts.