Chinese contentNeutronGym：面向 LLM 智能体的中子仪器设计评测环境
用中子仪器设计测智能体会不会物理，7个模型最多做对7/16题，RL把Qwen3-8B从11%练到77%，比32B还强。
Latest
Every published item, in order · 825 items
Oct 5
Chinese contentiS-KV：基于块增量 SVD 的在线 KV 缓存压缩方法
长链推理时 KV 缓存会爆炸，这篇提出用增量 SVD 压到 4-5 倍还能保住准确率，比删 token 的老办法稳，做推理部署的可以看看。
Chinese contentQwen3.8-Flash-Next 加 480KB 控制向量，本地部署减拒答版本
不想云端模型动不动拒答的话，这个用 480KB 控制向量改 Qwen3.8-Flash-Next 的法子挺有意思，12GB 显存就能跑。
Oct 4
Chinese content本地部署 Qwen3.8 千亿参数模型体验：192k 上下文够用
有人在低配电脑上跑通了千亿参数的 Qwen3.8，上下文 192k 也够日常用，想本地部署的可以看看他的实际感受
Chinese content开发者自研推理引擎 tqllm，单卡 RTX 3090 跑 Qwen 27B 达每秒 273 token
一张 RTX 3090 把 27B 模型跑到每秒 273 token，作者自己写的引擎，速度直接把普通部署方案甩开了。
Chinese contentStrata：12G显存本地跑Qwen3.8-Flash-Next 125B模型
12G显存的电脑也能跑125B的Qwen3.8-Flash-Next了，Strata把常用专家放显卡、全量放内存，项目已在Github开源。
Aleph Alpha's Kolibri: 78B open-weight model built for German
A European lab reaching benchmark parity with Chinese and American open models in a bilingual German-English model suggests small-model training capability has spread well beyond the usual hubs. The Apache 2.0 license makes it directly usable for on-premise deployments, which matters for German public-sector and enterprise buyers with data-sovereignty requirements. Teams comparing open models should weigh its German-language edge against its weaker SWE-Bench Verified result.
Microsoft paper ScholarEvolve evolves agent harnesses from research papers
The counterintuitive idea is that the literature, not runtime traces, is the highest-value source of harness improvements. Builders tuning agent scaffolds around a fixed model can replicate the module-and-topic-modeling loop with their own paper corpus instead of paying for a model upgrade. Results on AppWorld Challenge and Tau2-Bench Telecom give a concrete baseline for comparing it against failure-log-driven harness tuning.
Chinese contentAWS Strands Labs 开源决策模型 Strands Decider 2B，单次推理约 115 毫秒
AWS 开源了个 2B 小模型，专门替智能体做选择，115 毫秒出结果还带置信度，做路由和工具调用判断挺实用。
Sebastian Raschka's GRPO Tutorial Lifts Qwen3-0.6B to 47% on MATH-500
The practical detail that GRPO needs no reward model or critic — a SymPy verifier and group-relative ranking replace both — makes frontier-style reasoning training feasible on a single consumer GPU. Practitioners building small reasoning models can follow the five-step training loop directly, though they should watch for the overtraining collapse Raschka documents at 400 steps. The series is free on his YouTube playlist.
iPhone 17 Pro Max boosts MacBook Pro AI prefill speed by 44%
The pipeline shows that consumer devices can be daisy-chained to work around unified-memory ceilings on laptops, with gains concentrated in prefill rather than generation. Developers running Qwen3.8-27B-class models on 24GB machines can try layer-splitting across USB-C devices to cut long-context ingest times. Compared with cloud inference, this keeps everything local but requires custom software and offers no speedup below 64K for output tokens.
Chinese content开源项目 Strata 让 12GB 显存跑千亿参数 Qwen3.8-Flash-Next
12GB 显存就能跑千亿参数模型，Strata 靠 MoE 冷热分离把专家塞进内存，5070 上实测 94 token/s，家用显卡玩家可以折腾了
Chinese contentCloudflare 发布 Clef 决策模型，基于 Qwen 打造
Cloudflare 用 Qwen 造了个决策模型，输出只按 schema 来，flash 版每百万 token 才 0.09 美元，比 Jev 便宜不少。
Amazon launches Strands Decider 2B model
This model enables local AI decision-making without cloud dependency, addressing privacy concerns for sensitive applications. Developers can implement it for real-time decision tasks in edge computing environments, with performance competitive to larger models while requiring minimal resources.
Oct 3
Oct 2
ProVer method for agent RL credit assignment
The approach addresses a fundamental limitation in current RL methods where all trajectory tokens receive equal credit, making it impossible to identify decisive steps. Developers working on agent systems can implement this method to improve training efficiency, particularly when using smaller judge models, potentially reducing computational costs while maintaining performance gains.
webAI releases TwIL-LM3-Pro model
This model demonstrates efficient parameter utilization for specialized logical reasoning, making advanced AI capabilities accessible on consumer hardware. Developers working on formal verification systems or educational tools can leverage this lightweight model for local deployment without cloud dependencies.
Cloudflare launches Clef multimodal decision model
Clef's integration of visual capabilities with extended context creates new possibilities for complex business decision systems. Developers can now build multimodal applications that process both text and images simultaneously, with Cloudflare offering customization options through their new reinforcement learning product. The model's performance on Jev benchmarks positions it as a strong alternative to existing decision-focused AI systems.
IT之家Source
Multi-Token Prediction Heads Research
This research significantly reduces the computational cost of implementing MTP heads, making advanced inference optimization more accessible. The findings could enable faster deployment of large language models without requiring massive pretraining resources. Developers and researchers working on model optimization can implement these techniques to improve throughput with minimal computational overhead.
Qwen3.8-27B ARC-AGI results
The significant performance drop between ARC-AGI versions suggests challenges in maintaining reasoning capabilities across different task complexities. Developers working on reasoning models should examine Qwen's approach to template design, particularly the omission of medium reasoning guidance, which may impact token efficiency and performance consistency.
Oct 1
Chinese contentQwen3.8-Max 修复 PVE 服务器故障
Qwen3.8-Max 实际修复服务器案例，对比 Grok-4.7 和 DeepSeek-V4.1-Flash 表现更优。
Chinese contentSemifactual Credit-Augmented Policy Optimization
SCAPO算法通过半事实稳定性改进了令牌级信用分配，在数学和分布外基准上表现优异。
Cross-World On-Policy Distillation for VLMs
This research addresses a critical gap in VLM distillation where students can match teacher outputs without relying on the same visual evidence. The CW-OPD method provides a corrective signal that makes visual evidence reliance explicit rather than implicit. Developers working with smaller models can leverage this approach to achieve performance comparable to much larger models while maintaining better visual grounding.
Activation-Conditioned Self-Distillation
ACSD provides a novel approach to token-level supervision throughout long responses by leveraging activation contrasts, offering a more stable alternative to existing methods. Researchers and developers working on reasoning models can implement this technique to improve mathematical accuracy without requiring reference solutions, potentially enhancing model performance across diverse training datasets.
webAI releases TwIL-LM3-Pro model
This model represents the growing trend of efficient post-training techniques enabling powerful reasoning on consumer hardware. For developers and researchers working with limited resources, TwIL-LM3-Pro offers state-of-the-art reasoning capabilities that can run locally without cloud dependency, making advanced AI more accessible for edge computing applications.