搜索“RAG”

lf-Flow vs. Flow Matching (FM). Left: generation error (Fréchet distance) per modality, each normalized to FM = 100 (lower is better). Right: success rate on manipulation tasks averaged over four task groups through finetuning (higher is better). Capabilities & Early Evaluations As a result, FLUX 3 is capable of mixing modalities and generating images and video+audio jointly; both from pure text prompts as well as when providing input references such as images and video. We are highlighting a fe

匹配字段正文

a variety of benchmarks. In particular, we tested 3.5 Flash Cyber on the CyberGym benchmark, which evaluates AI agents against hundreds of real-world software vulnerabilities. Leveraging the low cost of 3.5 Flash Cyber by configuring CodeMender to call 3.5 Flash Cyber up to five times for a single, final report, the overall agent achieved competitive performance against significantly larger models on CyberGym . Competitor results are sourced from provider self-reported scores We also stress-test

匹配字段正文

ly $0.95 in and $4 out. K3 is also the most expensive LLM any Chinese AI company has ever released over the past two years, according to the model tracker of China AI Index. K3 averages about $0.94 per Intelligence Index task, close to GPT-5.6 Sol ($1.04) and roughly half the price of Opus 4.8 ($1.80). Yet some users on Xiaohongshu (Red Note) complained by using K3 they burned through their Kimi monthly usage far sooner than expected. One X user said K3 ends up being more expensive than Fable 5.

匹配字段正文

Beyond raw token volume, we invested heavily in data filtering and curation: deduplication, quality scoring, and domain-focused selection so that the data mixture stayed high-coverage and high-signal. We scaled reinforcement learning with a strong focus on per-token intelligence. Our RL training covers hundreds of thousands of tasks, centered on multi-step software engineering and other technical work, with automated and model-based grading. Our stack is built for highly asynchronous training,

匹配字段正文

izon asynchronous tasks, an important component of high-quality software engineering. SWE-1.7 is available today in Devin (Web, Desktop, and CLI) via Cerebras at 1000 TPS. We encourage you to try it for yourself! The rest of this post covers how we trained SWE-1.7: the infrastructure, algorithms, and data work behind our model. We cover four important components that stand out. Preserving entropy and stabilizing training: Long RL runs face two challenging problems: entropy collapse, and instabil

匹配字段正文

serving all of the learning signals. In practice, this method transforms training runs that would take months into runs that complete in days. Online reinforcement learning We leverage our knowledge of post-training LLMs at scale, using online reinforcement learning, to boost the performance of Robostral Navigate. After the supervised training stage, we further improve the model's performance using CISPO, an online reinforcement learning algorithm. This enables the model to learn from trial and

匹配字段正文

cient LFM2 architecture, LFM2.5-230M is considerably faster than similar-sized models, including SSM hybrids and Gated Delta Networks. On both a Raspberry Pi 5 and a Qualcomm Snapdragon Gen4 (Samsung Galaxy S25 Ultra), it delivers the highest prefill and decode throughput in its class while keeping the smallest memory footprint. We tune the flash-attention flag per device to maximize prefill on each platform: enabled (-fa 1) on the Raspberry Pi 5 and disabled (-fa 0) on the Snapdragon Gen4. GPU

匹配字段正文

or self-hosted deployments. It outperforms leading OCR systems in human evaluations and benchmarks while offering cost-efficient, high-throughput processing for enterprise search, RAG, and agentic workflows. Available via API or Document AI, it provides structured outputs for custom pipelines or no-code applications, with self-hosting options for data privacy. Today, we're releasing Mistral OCR 4, featuring bounding boxes, block classification, and inline confidence scores alongside extracted te

匹配字段正文

AgentWorld-397B-A17B achieves the highest overall score (58.71), outperforming GPT-5.4 (58.25) and other frontier models. Qwen-AgentWorld-397B-A17B achieves the highest overall average (58.71), surpassing GPT-5.4 (58.25) and all other frontier models. The advantage is most pronounced on Terminal and SWE, the two domains where predictions require accurate modeling of code execution state and tool API behavior. At the 35B-A3B scale, the three-stage pipeline lifts the overall average by +8.66 point

匹配字段正文

zon Manipulation We develop a VLM-driven agentic VLA system, in which the base Qwen-3.5 model serves as the high-level planner and Qwen-RobotManip handles low-level execution. Leveraging its capabilities in scene understanding, spatial reasoning, and task progress assessment, Qwen-3.5 decomposes a complex high-level instruction into a sequence of atomic subtasks, which are then executed by the VLA. This division of labor significantly improves robustness to OOD scenes and instructions. We illust

匹配字段正文

ce and is competitive with leading open-source TTS models on widely used open benchmarks such as seed-tts eval, as well as, on our own newly proposed ZTTS1-Eval benchmark. We encourage you to listen to the comparisons below. ZONOS2 performs especially well on speaker similarity and prosody metrics, a testament to the versatility and naturalness of its voice cloning capabilities. Continuing Zyphra’s tradition of architectural innovations, ZONOS2 is also the first open-source MoE text-to-speech mo

匹配字段正文

ends most of its time simply waiting for the next "keystroke." DiffusionGemma reverses this inefficiency. Instead of predicting words sequentially, it drafts an entire 256-token paragraph simultaneously. By giving the computer's processor a larger chunk of work at once, DiffusionGemma utilizes your hardware to its full potential. It upgrades your model inference from a single, sequential typewriter to a massive printing press that stamps the entire block of text simultaneously. DiffusionGemma te

匹配字段正文

l, Claude Opus 4.8. To release the model both safely and quickly, we’ve tuned these safeguards conservatively—they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions. With more capable models arriving in the coming months, we’re working to improve our safeguards and reduce false positives as quickly as we can. For a small group of cyberdefenders and infrastructure providers, we’re also launching Claude Mythos 5. It’s the same underlying model as Fa

匹配字段正文

uce verification overhead and increase concurrency, allowing high acceptance lengths to translate directly into high inference throughput: In the Coding scenario, we achieve an average acceptance length of 6.30, with some samples reaching a maximum of 7.14 — meaning 6–7 out of the 8 draft tokens per verification round are accepted. The draft model remains lightweight while pushing acceptance rates to levels that deliver real end-to-end gains. We also observe that in more semantically divergent,

匹配字段正文

apart, the gradients no longer reflect the data the model actually sampled, and training becomes off-policy and unstable. We track this with train rollout logprob abs diff, the average absolute difference between the log-probabilities assigned to each token by SGLang and Megatron. This value stays around 0.01 throughout the run, a small gap that suggests the rollout and training policies remain closely aligned and the pipeline is behaving on-policy Figure: train rollout logprob abs diff over a f

匹配字段正文

text-to-speech model we’ve built to date. It’s a significant leap from its predecessor across every dimension that matters to production voice experiences: fidelity, language coverage, speaker consistency, and emotional range. It is built for the products and services where voice quality directly impacts user experience: assistants or customer support that represent your brand, audiobooks that hold attention over hours, and accessibility experiences where voice is the only interface. It’s also

匹配字段正文

EURS benchmark using the ‘test’ split. FLEURS is a standard public multilingual dataset that we used internally for competitive evaluations. The model achieves state-of-the-art average WER across 43 languages and leads in 18 of them – outperforming GPT-4o-Transcribe, Scribe v2, and Gemini 3.1 Flash Lite. See more at Transcribe 1.5 model page; See more in the model card, which includes pricing details. MAI-Transcribe-1.5 is a batch transcription model, and comparisons are made against similar sys

匹配字段正文

wn on FLEURS multilingual transcription benchmark, and 3 on the Artificial Analysis leaderboard; Leading accuracy x speed on Artificial Analysis leaderboard; Expanded language coverage from 25 to 43; Can transcribe an hour of audio in under 15 seconds. Up to five times faster on long audio than Gemini 3.1, Scribe v2, GPT-4o-Transcribe; Includes Keyword Biasing, enabling the model to be aware of domain specific terminology which improves WER by up to 30% on FLEURS; Optimized for real-world use ca

匹配字段正文

-Bench Pro, SWE-Bench Multilingual, and Terminal Bench 2 using the same production harness that developers use for their everyday coding tasks. We measured task success and the average number of solution tokens required to complete each task. MAI-Code-1-Flash outperforms Claude Haiku 4.5 across all core coding benchmarks tested, with higher pass rates on all 4 evaluations, including a +16-point lead on the diverse, real-world tasks of SWE-Bench Pro (51.2% vs. 35.2%). It’s not just smarter; it’s

匹配字段正文

and that includes models 40–50% larger, and the 7B model is consistently competitive with the strongest 7–8B open-weight VLMs across nearly every category. We view this as an encouraging sign for scaling hybrid models further. Inference Efficiency Zamba2-VL’s principal advantage is in inference. The Transformer's quadratic attention scales poorly to the long token streams that arise in the multimodal regime. A single high-resolution image can contribute several thousand vision tokens, and a shor

匹配字段正文