Mistral

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：The French AI lab extends its push for regional control of AI from Europe to the Middle East.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：According to OpenRouter data, agentic AI workloads consume 15x more tokens than a simple chat request. Why? Consider what happens when an AI agent researches a company for an investment decision. The agent queries financial databases, searches news and filings, invokes a sub-agent to run peer comparisons and model valuations, then synthesizes everything into a […]

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2608.18090v1 Announce Type: new Abstract: Inside a modern language model sits a single internal direction that tracks how positive or negative a sentence feels. We show how to find this valence axis (V-axis) from just 9 emotion category names plus 50 short narrative paragraphs per emotion -- about 1,500 fewer labels than the usual supervised approach -- and that the same direction appears in vision, audio, and human-brain encoders never jointly trained. The recipe: embed nine emotion-anchored story sets in a frozen encoder, take the top principal direction of the nine averaged embeddings. Projecting new inputs onto it captures 93% of supervised performance on SST-2 (Llama-3-8B-Instruct, AUC 0.772 vs. 0.828), correlates with human valence ratings on 11,811 EmoSet images at r=0.636, reaches AUC 0.906 on ESC-50 audio (p12). A 2-parameter classifier trained on text labels transfers to images (AUC 0.961), audio (0.764), and brain recordings (0.828) without target-modality labels; a generic 16-D subspace stays at chance (0.525). The recipe is bounded to continuous attributes -- seven tests on categorical concepts return near-chance -- and steering is family-specific (Llama/Mistral yes, Qwen/Gemma no).

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2608.18089v1 Announce Type: new Abstract: Instruction-tuned models often refuse harmful requests in English but comply with the same requests in Yoruba, Igbo, Igala, and Hausa. This suggests that the refusal mechanism is present in the residual stream but fails to activate for low-resource inputs. Recovering it normally requires labelled target-language data and retraining, neither of which is available at scale for most African languages. We introduce Latent Space Refusal Anchoring (LSR-Anchoring), a training-free method that extracts the refusal direction from English prompts and clamps it onto the residual stream at inference time. The primary variant, Mean-Activation Steering (MAS), operates across the four architectures we tested: Llama-3-8B, Llama-3.1-70B, Mistral-7B-Instruct, and Qwen2.5-7B. On Mistral and Qwen it recovers safety with benign degradation below 0.08. On Llama-3-8B it overcorrects, with Degraded Performance on Legitimate prompts (DPL) reaching 1.00. We address this with SAE-Derived Steering (SDS), which replaces the dense mean-difference direction with a single Sparse Autoencoder (SAE) feature and reduces Kullback-Leibler (KL) divergence by 3.5-7x without benign collapse. Four languages transfer positively, but Arabic fails on every architecture and at every steering magnitude, indicating a geometric mismatch rather than a baseline effect. Massive Multitask Language Understanding (MMLU) accuracy drops remain below 0.35 percentage points at every effective steering magnitude.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Mistral AI wants to turn European AI sovereignty from a talking point into a product — one with a service-level agreement attached. The French artificial intelligence company announced Tuesday a three-part expansion of…

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2608.11210v1 Announce Type: new Abstract: Bayesian calibration of process-based models requires a prior distribution for each model parameter. Despite decades of methodological work, researchers almost always fall back on uniform priors. The main reason is that building informative priors from scientific literature is slow and needs both domain and statistical expertise. We present \textbf{Distribird}, an agentic web application that automates this process. Given a parameter name, physical description, and domain context, Distribird deploys a multi-agent pipeline that searches the literature, extracts and weights reported values by domain relevance, and fits a probability distribution via AIC model selection. When no literature is available, the system falls back to sensible uninformative alternatives, and clearly reports both the evidence behind and the confidence level of every prior it produces. It is designed for the problems where the models have physically interpretable parameters, where domain knowledge exists in the published literature. We evaluate the tool on 24~parameters across 10 scientific domains comparing three open-weight models (Qwen3.6 27B, Gemma 4 31B, Mistral Small 4 119B) with a single-prompt LLM baseline. On prior quality the full pipeline \emph{matches} this baseline. Every prior is traced to the specific papers and values from which it was constructed; a built-in validity layer declines to produce priors for out-of-scope requests, whereas the single-prompt baseline returns confident but unfounded priors for them in 11 of 30~model--parameter cases; and every language-model call runs locally, so no parameter description or unpublished modelling detail is transmitted to a third-party LLM provider (only generated search terms reach the public literature databases). For scientific use, we argue these properties matter more than a marginal improvement in point-estimate accuracy.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：The Paris-based vendor continues to build European AI infrastructure.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：That’s a lot of people chatting with their AI friends all day. | Image: Google For the 14th time, a Google product has hit 1 billion users. Google CEO Sundar Pichai posted on X that a billion people are using Gemini every month, and that Gemini is Google's fastest-growing product ever. A billion users is a huge milestone, but Google isn't the first AI app to hit it. OpenAI's ChatGPT hit the mark a few weeks ago, though the company buried the announcement that "more than 1 billion people are putting ChatGPT to work" in an otherwise anodyne blog post about how people use AI. External data suggested ChatGPT crossed 1 billion users as early as this June, but OpenAI hadn't announced anything until that post on August 6th. … Read the full story at The Verge.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：Mistral AI has released Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier that frames content moderation as a single yes/no question instead of a fixed harm taxonomy. Operators supply the policy as a plain-language query at inference time and get back a calibrated safety score from one forward pass — no retraining required to re-target the model. Built on Ministral-3-3B-Base-2512 with a Pixtral vision encoder and trained on roughly 54.1M samples, it reports 84.9% average F1 on text safety (matching GPT-OSS-Safeguard-20B), 83.8% on multimodal safety, and 91.3% on Mistral's adaptability benchmark — while fitting in 16GB of VRAM under an Apache 2.0 license. The post Mistral AI Releases Shieldstral 1.0 3B: An Open-Weights Policy-Adaptive Multimodal Safety Classifier Matching Models 7× Its Size appeared first on MarkTechPost.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2608.05162v1 Announce Type: new Abstract: Pooling is a consequential but under-examined design choice in decoder-only concept representation work: practitioners must collapse token-level hidden states into a passage-level vector, yet no shared protocol exists for comparing this choice across concepts, models, and tasks. Reported gains are confounded by simultaneous changes in dataset, layer, construction method, and pooling rule, making principled decisions impossible. We introduce PoolBench, a benchmark that isolates pooling as the experimental variable under a fixed evaluation protocol. PoolBench covers 17 concepts, 19 pooling strategies, and 3 open-weight decoder-only models (Llama-3.1-8B, Gemma-2-9B, Mistral-7B), evaluated on a single audited corpus of 37,693 real-text passages. The primary axis is linear separability (D1/AUROC); steered concept prevalence (D2/SCP) and output-level disentanglement (D3) serve as diagnostic axes. The primary finding is decisive: W4_hierarchical reaches a cross-model mean AUROC of 0.7799, while the widely adopted P1_last_token baseline reaches only 0.7640 and is statistically significantly worse (Friedman+Nemenyi, p = 2.0e-36; 77 significant pairs among 18 effective strategies). Rankings are stable across layers (rho = 0.961--0.990). A key negative result: strong detection does not imply strong steering -- D2 and D3 are substantially weaker than D1 for most concepts, indicating a fundamental representational limit rather than a pooling failure. On mid-difficulty concepts, W4_hierarchical outperforms P1_last_token by 0.042--0.113 AUROC; construction method choice (DiffMean vs. REPE) has a larger effect (delta AUROC 0.15) than pooling (delta AUROC 0.016), establishing the correct practical hierarchy. We release the corpus, pre-extracted activations, scorer models, steering vectors, and evaluation code as a reusable protocol for pooling research.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：arXiv:2608.04130v1 Announce Type: new Abstract: Vision-language models for autonomous driving primarily rely on cameras and LiDAR, leaving 4D radar largely unexplored as a standalone perceptual modality despite its robustness to adverse visibility and direct measurement of radial velocity. We introduce Radar4D-VLM, a radar-only temporal vision-language model that reasons from ten consecutive 4D-radar point-cloud sweeps without camera or LiDAR input. Radar4D-VLM extracts geometrically grounded object proposals and organizes radar evidence into a compact hierarchy of object, scene, and kinematic tokens. A parameter-efficient projector maps these tokens into frozen language backbones, while auditable prediction heads jointly model object count, spatial distribution, motion state, collision risk, semantic category, and radial velocity. Radar4D-VLM combines proposal-grounded temporal object tokenization, global scene context, and explicit kinematic tokens within a unified frozen-backbone interface. On sequence-isolated K-Radar development validation, its Top-64 proposal recall reaches 98.13% at 4 m, exceeding fixed-lattice and uniform-random controls by 6.40 and 22.83 percentage points, respectively. We further evaluate 24 matched runs spanning eight frozen Qwen, Phi, Mistral, Llama, and Gemma backbones under an identical adaptation budget. The radar-token interface remains compatible across all five language-model families, while matched aligned, permuted, and no-language controls show sensor dependence but no stable direct-head gain from aligned language supervision. These results establish a reproducible foundation for radar-only multimodal scene and motion reasoning while separating interface compatibility from the benefit of language supervision.

AI 服務暫時不可用，以下為來源摘要，待恢復後補全翻譯：<p>I released <a href="https://llm.datasette.io/en/stable/changelog.html#v0-32">LLM 0.32</a> this morning, the most significant new version of LLM since the initial launch of the project. The new version includes support for visible reasoning traces, server-side provider tools, redesigned content-addressable SQLite logs, new models, and new features enabled by the OpenAI Responses API. I also released new versions of the <code>llm-anthropic</code>, <code>llm-gemini</code>, and <code>llm-openrouter</code> plugins, each with substantial updates of their own.</p> <h4 id="headline-features-for-llm-cli-users">Headline features for LLM CLI users</h4> <p>Running LLM against reasoning models now <strong>displays their reasoning traces</strong> to standard error, so you can see what they are "thinking" without that information being included in the standard output that you might pipe to another tool. Add <code>-R/--hide-reasoning</code> to turn this off.</p> <p><img src="https://static.simonwillison.net/static/2026/best-pelicans.gif" alt="Running llm &quot;think about the best thing about pelicans&quot; in the macOS terminal window - grey text outputs saying Exploring pelican qualities, then after a paragraph of that a white paragraph of text comes out saying: The best thing about pelicans is their wonderfully oversized, practical design: that enormous bill and pouch look comical, but they make pelicans remarkably skilled fishers. Even better, many species cooperate—working together to herd fish before scooping them up. They’re a great mix of goofy, graceful, and surprisingly clever." style="max-width: 100%;" /></p> <p>LLM includes support out-of-the-box for the <strong>GPT-5.6 model family</strong>, and the new default model used with <code>llm "prompt"</code> is now the inexpensive but capable <strong>GPT-5.6 Luna</strong>.</p> <p>LLM calls can now use <strong>server-side tools</strong> from various providers. OpenAI provide <a href="https://llm.datasette.io/en/stable/openai-models.html#code-interpreter">a code execution environment</a> as a server-side tool; LLM can now run prompts that benefit from that like so:</p> <div class="highlight highlight-source-shell"><pre>llm --tool CodeInterpreter <span class="pl-s"><span class="pl-pds">'</span>Show current python and SQLite versions<span class="pl-pds">'</span></span></pre></div> <p>OpenAI also gets a <a href="https://llm.datasette.io/en/stable/openai-models.html#web-search">WebSearch</a> tool.</p> <p>The <a href="https://github.com/simonw/llm-anthropic">llm-anthropic</a> plugin adds <a href="https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search">WebSearch</a>, <a href="https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch">WebFetch</a>, <a href="https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution">CodeExecution</a>, and <a href="https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector">AnthropicMCP</a>, which looks like this:</p> <div class="highlight highlight-source-shell"><pre>llm -m claude-sonnet-5 -T <span class="pl-s"><span class="pl-pds">'</span>AnthropicMCP("https://datasette.simonwillison.net/-/mcp")<span class="pl-pds">'</span></span> \ <span class="pl-s"><span class="pl-pds">'</span>how many rows in the blog_blogmark table?<span class="pl-pds">'</span></span></pre></div> <p>That causes Anthropic to execute MCP calls against my new <a href="https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp">datasette-mcp</a> plugin as part of a single request/response interaction with their API.</p> <p>The new <strong>llm openai endpoint</strong> command provides a tool for <a href="https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it">executing prompts against <em>any</em> OpenAI compatible endpoint</a> as a one-liner. These aren't logged, which makes this a handy tool for running one-off prompts against anything that speaks the lingua franca of the LLM API world.</p> <p>Here's how I use that to run prompts against Gemma 4 12B running in my localhost <a href="https://lmstudio.ai">LM Studio</a> API, via <code>uvx</code> (no LLM installation required) and mixing in the <a href="https://github.com/simonw/llm-tools-quickjs">llm-tools-quickjs</a> tool plugin for good measure:</p> <div class="highlight highlight-source-shell"><pre>uvx --with llm-tools-quickjs \ llm openai endpoint http://localhost:1234/v1 -m google/gemma-4-12b \ -T QuickJS <span class="pl-s"><span class="pl-pds">'</span>Use QuickJS to multiply 3434 * 2434<span class="pl-pds">'</span></span> --td</pre></div> <p><img src="https://static.simonwillison.net/static/2026/openai-endpoint-gemma.webp" alt="Output reads Tool call: QuickJS_execute_javascript({'javascript': '3434 * 2434'}) 8358356 The result of 3434 * 2434 is 8,358,356." style="max-width: 100%;" /></p> <h4 id="new-features-in-the-python-api">New features in the Python API</h4> <p>LLM's Python API previously required you to create a conversation and then send messages to it one at a time. This was an abstraction over the true nature of LLMs, where each request carries a complete history of the messages that came before it. That abstraction started to get in the way for some more advanced cases, so the new release introduces a <code>model.prompt(messages=[])</code> parameter that can be used like this:</p> <pre><span class="pl-k">import</span> <span class="pl-s1">llm</span> <span class="pl-k">from</span> <span class="pl-s1">llm</span> <span class="pl-k">import</span> <span class="pl-s1">user</span>, <span class="pl-s1">assistant</span>, <span class="pl-s1">system</span> <span class="pl-s1">model</span> <span class="pl-c1">=</span> <span class="pl-s1">llm</span>.<span class="pl-c1">get_model</span>(<span class="pl-s">"gpt-5.6-luna"</span>) <span class="pl-s1">response</span> <span class="pl-c1">=</span> <span class="pl-s1">model</span>.<span class="pl-c1">prompt</span>(<span class="pl-s1">messages</span><span class="pl-c1">=</span>[ <span class="pl-en">system</span>(<span class="pl-s">"You are a helpful pirate."</span>), <span class="pl-en">user</span>(<span class="pl-s">"What is the capital of France?"</span>), <span class="pl-en">assistant</span>(<span class="pl-s">"Paris, matey."</span>), <span class="pl-en">user</span>(<span class="pl-s">"And Germany?"</span>), ]) <span class="pl-en">print</span>(<span class="pl-s1">response</span>.<span class="pl-c1">text</span>())</pre> <p>LLM previously returned an iterable sequence of strings from each prompt. This worked great when models returned a string response, but failed to predict the weird shape that models would evolve towards. Today many models return a mix of reasoning text, output strings, tool calls, and even image attachments. With LLM 0.32 you can <a href="https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events">do this instead</a>:</p> <pre><span class="pl-k">for</span> <span class="pl-s1">event</span> <span class="pl-c1">in</span> <span class="pl-s1">model</span>.<span class="pl-c1">prompt</span>(<span class="pl-s">"Explain cats"</span>).<span class="pl-c1">stream_events</span>(): <span class="pl-k">if</span> <span class="pl-s1">event</span>.<span class="pl-c1">type</span> <span class="pl-c1">==</span> <span class="pl-s">"reasoning"</span>: <span class="pl-en">print</span>(<span class="pl-s">f"[thinking] <span class="pl-s1"><span class="pl-kos">{</span><span class="pl-s1">event</span>.<span class="pl-c1">chunk</span><span class="pl-kos">}</span></span>"</span>, <span class="pl-s1">end</span><span class="pl-c1">=</span><span class="pl-s">""</span>, <span class="pl-s1">flush</span><span class="pl-c1">=</span><span class="pl-c1">True</span>) <span class="pl-k">elif</span> <span class="pl-s1">event</span>.<span class="pl-c1">type</span> <span class="pl-c1">==</span> <span class="pl-s">"text"</span>: <span class="pl-en">print</span>(<span class="pl-s1">event</span>.<span class="pl-c1">chunk</span>, <span class="pl-s1">end</span><span class="pl-c1">=</span><span class="pl-s">""</span>, <span class="pl-s1">flush</span><span class="pl-c1">=</span><span class="pl-c1">True</span>) <span class="pl-k">else</span>: <span class="pl-en">print</span>(<span class="pl-s">f"Other event: <span class="pl-s1"><span class="pl-kos">{</span><span class="pl-s1">event</span><span class="pl-kos">}</span></span>"</span>)</pre> <p>Combine these features and we can <em>finally</em> provide a robust implementation of the semi-standard OpenAI chat completions API, which I've now released as the <a href="https://github.com/simonw/llm-chat-completions-server">llm-chat-completions-server</a> plugin:</p> <div class="highlight highlight-source-shell"><pre>llm install llm-chat-completions-server llm chat-completions-server --port 9000 <span class="pl-c"><span class="pl-c">#</span> Server is now running on http://127.0.0.1:9000/v1</span></pre></div> <p>Now you can run prompts against LLM via that server, using the new <code>llm openai endpoint</code> command!</p> <div class="highlight highlight-source-shell"><pre>llm openai endpoint http://127.0.0.1:9000/v1 <span class="pl-s"><span class="pl-pds">'</span>hello<span class="pl-pds">'</span></span> -m gpt-5.4-mini</pre></div> <p>The bigger challenge with that kind of API concerns logging. If we're going to support the pattern where the message sequence is appended to on every request, ideally we can avoid logging all of that duplicate JSON for every turn.</p> <p>The solution is the new <a href="https://llm.datasette.io/en/stable/logging.html#the-message-store">content-addressable message store</a>, modeled after Git. You can see the new schema for that <a href="https://llm.datasette.io/en/stable/logging.html#sql-schema">in the documentation</a>, but the <code>llm logs</code> and <code>llm logs --json</code> commands have both been upgraded to convert that format back into something that's easy to consume.</p> <h4 id="and-the-rest">And the rest</h4> <p>There is a whole lot more in this release. The <a href="https://llm.datasette.io/en/stable/changelog.html#v0-32">0.32 release notes</a> are pretty comprehensive, and the notes for <a href="https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30">0.32rc2</a>, <a href="https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30">0.32rc</a>, <a href="https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09">0.32a3</a>, <a href="https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12">0.32a2</a>, and <a href="https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28">0.32a0</a> should fill in any gaps.</p> <p>Existing LLM plugins should all continue to work, but plugins that provide extra models will need to be upgraded to 0.32 in order to participate fully in the new streaming events system. There's a guide to implementing plugins with <a href="https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events">Structured messages and streaming events</a> in the documentation.</p> <p>I've updated some of my own plugins:</p> <ul> <li> <a href="https://github.com/simonw/llm-anthropic/releases/tag/0.26">llm-anthropic 0.26</a> adds support for the Claude 5 family of models, plus <code>WebSearch</code>, <code>WebFetch</code>, <code>CodeExecution</code>, and <code>AnthropicMCP</code> server-side tools.</li> <li> <a href="https://github.com/simonw/llm-gemini">llm-gemini</a> and <a href="https://github.com/simonw/llm-openrouter">llm-openrouter</a> and <a href="https://github.com/simonw/llm-mistral">llm-mistral</a> are nearly there, releases coming soon.</li> </ul> <h4 id="i-guess-llm-is-an-agent-framework-now">I guess LLM is an agent framework now</h4> <p>Quite a few of the lower-level tools changes in this release were driven by the needs of <a href="https://agent.datasette.io/">Datasette Agent</a>. When I started work on LLM, the term "agent" had such a vague definition that I refused to use it. In <a href="https://simonwillison.net/2025/Sep/18/agents/">September 2025</a> I came around to the idea that "<strong>An LLM agent runs tools in a loop to achieve a goal</strong>" is well established enough now that I could stop avoiding the term entirely.</p> <p>Tool chains can now <a href="https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause">pause for human approval</a> and <a href="https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume">resume from a stored message history</a> - both needed by Datasette Agent.</p> <p>Looking at LLM today it's beginning to look very agent-shaped to me. There's something neat about having a CLI utility that can mix and match different tools from different sources with different models all as a one-liner, and that includes a Python library powerful enough to build systems like <a href="https://agent.datasette.io/">Datasette Agent</a> and <a href="https://github.com/simonw/llm-coding-agent">llm-coding-agent</a>.</p> <p>Maybe the next version of LLM will bake the concept of an "agent" into the core library. I'm still trying to figure out what that would look like.</p> <p>Tags: <a href="https://simonwillison.net/tags/projects">projects</a>, <a href="https://simonwillison.net/tags/releases">releases</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/openai">openai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/anthropic">anthropic</a></p>

AI模型中心是一個集中平台，幫助用户對比來自Meta、阿里巴巴、谷歌、Mistral等領先開發者的開源大語言模型。它提供了超過100個活躍模型的詳細規格，包括上下文窗口、架構、參數規模、許可證和基準測試結果。

一項新研究評估了在NVIDIA H100 GPU上啓用機密計算對大型語言模型推理性能的影響。測試使用Mistral-7B和Qwen3-30B-A3B模型，發現機密模式使首令牌延遲平均增加21.8%-27.8%，全局令牌吞吐量下降17.7%-21.1%，且較大模型更早達到飽和。結果表明機密GPU推理在負載下仍可保持可用吞吐量，但容量規劃需考慮性能損失和早期飽和現象。

NVIDIA Vera Rubin NVL72 正加速生產，與 CoreWeave、Google Cloud、Microsoft Azure 和 Oracle Cloud Infrastructure 等合作伙伴共同部署。該平台通過極致協同設計實現最高的每瓦性能和最低的令牌成本，在 DeepSeek-R1 基準測試中每兆瓦吞吐量比 Grace Blackwell NVL72 提升 10 倍。Vera Rubin 還支持歐洲開放模型時代，與微軟和 Mistral 合作擴展 AI 基礎設施。

本文對比了六款適合單張24GB GPU（如RTX 3090/4090）的開放權重模型，涵蓋Qwen3.6、Gemma 4、Mistral Small等，並解釋了內存分配、量化策略以及各模型的優勢場景。

本文對比了四種主流的AI編程代理：Mistral Vibe for Code、Claude Code、Cursor和OpenAI Codex，針對從功能腳手架到拉取請求的完整工作流進行評分。Mistral Vibe以22/25的總分領先，憑藉成本、開放性和控制力獲勝；Claude Code和Codex並列21/25；Cursor得16/25。文章詳細分析了每個工具在腳手架、測試循環、PR及異步工作流、覆蓋範圍、成本與開放性五個維度的表現。

Mistral AI 推出了一款新型視覺模型，機器人僅需一個RGB攝像頭和自然語言指令即可在陌生環境中導航。

Mistral AI 推出了 Robostral Navigate，一個 8B 參數的具身導航模型。該模型僅使用單個 RGB 攝像頭，無需 LiDAR 或深度傳感器，即可根據自然語言指令驅動機器人。在 R2R-CE 驗證未見過的場景中，它達到了 76.6% 的成功率，這得益於其指向方法、前綴緩存訓練和 CISPO 在線強化學習。

本研究探索利用機器學習自動對大型文學語料庫進行主題索引，以伏爾泰作品為案例，比較了多種模型，其中Mistral系列4位量化模型F1得分達0.67，證明了自動索引的潛力。

本文介紹了Director，一種新的分佈式MoE推理系統，通過預測驅動的在線專家放置優化，顯著降低端到端延遲。系統採用輕量級級聯預測器或低比特量化副本預測專家激活模式，結合近乎零停機的在線遷移模塊，以及基於鬆弛優化的專家放置算法，在多項式時間內達到(1+ε)近似比。實驗表明，在Mistral、DeepSeek和Qwen等流行MoE模型上，相比現有工作延遲降低11%~55%。

作者從個人編碼和審計經驗出發，對2026年中的主流AI模型進行非正式分級，涵蓋Anthropic Fable、OpenAI Sol、Mistral、Gemini和DeepSeek等模型，並融入美國出口管制和歐洲視角的評論。

AnyFile Translator 是一款AI翻譯助手，可在Google Chat中直接翻譯文件、網頁鏈接和文本，保留原始佈局和格式，支持超過100種語言。它還具備AI寫作功能，可生成並翻譯內容。適合國際團隊和全球客户使用。

本文詳細介紹瞭如何使用 Amazon Bedrock AgentCore 和 Mistral AI Studio 構建並連接一個生產就緒的電子商務 MCP（模型上下文協議）服務器。內容涵蓋 MCP 工具實現、雙層 JWT 認證、AWS CDK 部署、與 Mistral AI Vibe 集成，以及使用 DynamoDB 和 Cognito 管理數據與身份的最佳實踐。

該論文對KIVI、TurboQuant、SnapKV和CaM等KV緩存優化技術進行了工作量感知的基準測試，評估了它們在Llama-3.1-8B-Instruct和Mistral-7B-Instruct-v0.3模型上的多文檔問答、單文檔問答、少樣本學習和摘要任務中的表現。結果表明，壓縮率本身並不能很好地預測端到端性能。KIVI4提供最穩定的質量，SnapKV在長上下文吞吐量方面表現最佳，而CaM在特定問答任務上取得顯著提升，但對工作負載敏感。該研究強調了根據工作負載選擇KV緩存機制的必要性。

Mistral AI 發佈了 Leanstral 1.5，這是一個基於 Apache-2.0 許可的 Lean 4 代碼代理模型。該模型採用 119B 混合專家架構，每令牌激活 6.5B 參數，上下文長度 256k。它在 miniF2F 上達到 100% 準確率，解決了 PutnamBench 中 587/672 的問題，並在 FATE-H 和 FATE-X 基準測試上實現了新 SOTA。此外，它還能發現真實軟件缺陷，已在 57 個開源倉庫中識別出 5 個未報告的錯誤。

Wiola是一種全新的小型語言模型架構，從基本原理設計，與GPT、LLaMA、Mistral或Falcon等現有模型無結構關聯。它引入了五種獨立創新的組件：螺旋旋轉位置編碼（SRPE）、門控跨層注意力（GCLA）、自適應令牌合併（ATM）、雙流前饋（DSFF）和WiolaRMSNorm歸一化。模型提供四種規模（120M、360M、700M和1.5B參數），完全兼容HuggingFace Transformers生態系統。

本文對Beckmann & Butlin (2026)關於LLM個體化的本體論框架提出質疑，認為其繼承了未論證的跨體制共指假設。通過Qwen3-4B-Instruct和Mistral-7B-Instruct-v0.2上的個性拓撲實驗，作者展示了四個經驗性楔子，共同削弱該假設，並提出體制索引個體化：表徵內容的身份單位是（載體，體制）對，而非僅載體。

本文形式化了基於Huber污染模型的LLM陪審團，並證明即使只有一個評委以LLM典型方式（模式崩潰、諂媚、安全拒絕）產生偏差，任何正污染都會導致PoLL產生無界偏差。通過將陪審團共識視為經典魯棒均值估計，作者提出RoPoLL，用幾何中位數替換聚合函數，實現了最優有限樣本崩潰點1/2。實驗表明，在13個開源評委（4B-675B）、三個獎勵模型基準和四種腐敗機制（高達50%）下，RoPoLL在每一種有偏腐敗類型上都優於PoLL：在匹配計算量的跨維度攻擊上提升約19%，在重尾拜占庭對手上提升數個數量級。一個38B參數的3評委RoPoLL委員會在30%雙模隨機腐敗下，在HelpSteer-2上以18倍參數優勢超越Mistral-Large-3（675B）1.31倍。

Bored People Chat 是一個極簡的匿名全球聊天室，無需註冊、無廣告、無機器人，旨在提供一個安全的空間，讓人們與陌生人交流。創建者受舊式聊天室和MMORPG的啓發，強調安全與AI輔助的審核。

LangChain 發佈了四個新的測試環境，用於評估大型語言模型（LLM）使用工具完成任務的能力，涉及規劃、函數調用和推理等關鍵技能。測試比較了 GPT-4、Claude 2.1、GPT-3.5 以及開源模型（如 Mistral 7b）的表現。關鍵發現包括：GPT-4 在關係數據任務中表現最佳，但在長時間軌跡中易出錯；Claude 2.1 在三個任務中與 GPT-4 相當；開源模型在多次函數組合上表現不佳；規劃能力仍是 LLM 的難點。

法國初創公司 Mistral AI 的 OCR 4 模型包括邊界框等功能，幫助用户更好地理解非結構化數據。

Mistral AI 發佈了 OCR 4，從純文本提取轉向結構化文檔輸出。每個塊返回邊界框、類型分類以及每頁和每詞的置信度分數。該模型支持170種語言，可在單個自託管容器中運行，並通過一個API端點將可引用的輸入提供給RAG、代理和企業搜索管道。

Mistral OCR 4 引入了邊界框、塊分類和置信度分數，在人類偏好測試中優於所有競爭對手，平均勝率達72%。支持170種語言，可單容器自託管，定價為每千頁4美元。

Mistral AI 宣佈今年夏天將推出一款新模型，該模型將開啓一個更大但稀疏的模型家族。7月將面向研究、政府等關鍵合作伙伴開放早期訪問計劃。

新聞驗證機構NewsGuard的審計發現，Mistral AI的聊天機器人Le Chat在有關伊朗戰爭的虛假聲明中，英文回覆有50%重複虛假信息，法語回覆有56.6%重複虛假信息。法國武裝部隊部使用的是定製版的Le Chat Enterprise，而非免費的消費者版本。

Mistral 發佈全新 AI 代理 Vibe，支持長期多步驟工作，包括郵件日曆管理、深入研究、文檔編寫以及編碼任務。Vibe 運行於旗艦 Mistral 模型，提供工作模式和代碼模式，並推出 VS Code 擴展和 CLI 更新。

基輔的Abrams軍事商店在戰爭期間迅速擴張，部分原因是標準軍用服裝的不足。許多訪客尋求的不僅僅是服裝。

上週末，應華盛頓要求，Anthropic突然下線了其最新、最強大的AI模型。這家美國公司稱，在白宮要求封鎖所有外國國民（包括其員工）的訪問後，它別無選擇。在國外，這一事件清醒地提醒人們，美國不僅主導着前沿AI，其政府還掌控着誰可以使用它的權力。特朗普政府的行動迅速、全面，且幾乎沒有預警或解釋。Fable 5和Mythos 5模型的空前關閉——它們本已受到限制在“高風險領域”使用的保障措施——為長期以來的論點注入了新的力量，即警告不要依賴美國的關鍵技術。這為已經在主張自身需要領導該技術的政治家、政府和企業提供了新的彈藥。在英國，AI和在線安全部長Kanishka Narayan沒有直接提及Anthropic、特朗普或美國，但利用這次關閉來論證英國必須發展自己的AI能力，並將其框定為國家安全問題。他説：“我們嚴肅對待每一個對我們主權構成的威脅，但我們還沒有學會以同樣的方式對待這個。”他稱AI是“我們時代的核心政治問題”，並認為英國必須決定這項技術將如何塑造其經濟、安全和主權，“否則別人會替我們決定答案。”在法國，反應更加直接，也更明確地指名美國。前總理Gabriel Attal稱這次關閉是“AI戰爭”的開始，並表示這表明法國如果依賴他人的關鍵技術就會脆弱。他將Anthropic模型的撤出比作伊朗封鎖霍爾木茲海峽，認為獲得AI現在是一個戰略瓶頸，法國必須為此做好準備。這並非全新論點。歐洲多年來一直擔心對美國的依賴，歐盟越來越強調在芯片、雲計算和AI等領域減少對外部提供商的依賴。但Anthropic關閉事件使事情變得更加緊迫，加劇了在特朗普領導下對美國作為盟友可靠性的深深不安——從貿易爭端到威脅退出北約。加拿大也得出了類似結論。總理Mark Carney表示，這一情況凸顯了僅依賴一個合作伙伴獲取AI等關鍵資源的風險。他説：“我們目前與Mythos和Fable所處的困境，就是過度依賴某些模型可能發生的情況。這個情況下沒有人做錯什麼。但如果我們只是接受這一點，不吸取教訓，不建立和多樣化，那麼我們就會做錯。”其他國家早已走上這條路。北京長期以來一直支持國內AI公司，中國是少數擁有能與美國前沿AI實驗室產品相媲美的模型的地方之一。但大多數政府和企業的規模和資源無法與美國或中國的前沿實驗室相比。主權AI並不總是意味着構建最大或最強大的工具。法國的Mistral和加拿大的Cohere表明，即使模型無法並駕齊驅，這些國家之外也能做出紮實的努力。其他國家，如新加坡和阿聯酋，專注於更狹窄但具有戰略意義的優先事項，如基礎設施或對本地語言更好的模型。當然，還有開源模型，其能力可能有一天會達到Mythos的水平，且難以被任何一方控制。特朗普可能將限制Mythos和Fable視為國家安全問題。但這一論點也是雙向的：既然華盛頓在問AI是否太過重要以至於不能讓人人都使用，其他政府也在問他們是否能承受由華盛頓來決定誰可以使用。Anthropic可能很快會重新上線Mythos和Fable。但恢復全球對美國AI的信任則是另一回事。無論關閉持續多久，它都揭示了訪問美國前沿AI模型的脆弱性。許多政府和企業不喜歡他們看到的——他們決心確保這種事不再發生。

Cohere發佈其首個開源編程模型North Mini Code（Apache 2.0許可），瞄準希望擁有和控制AI基礎設施的開發者。該30B MoE模型可在單張H100 GPU上運行，在代理編程任務上與Mistral、Qwen和Gemma競爭。

Tokens 4 Breakfast 是一款 macOS 菜單欄應用，用於實時追蹤和監控AI工具（如 Claude、OpenAI、Cursor、Copilot、Gemini、DeepSeek、Mistral 等）的令牌使用、訂閲費用和速率限制，幫助開發者避免意外超支。應用提供免費版（支持單一供應商）和一次性付費的 Pro 版（$7.99），所有數據本地處理，無需登錄或雲服務。

法國人工智能初創公司Mistral AI正在談判新一輪約30億歐元的融資，估值約為200億歐元，旨在加強其在歐洲AI領域的地位。

本文介紹如何通過Ollama和Scikit-LLM Python庫，免費使用本地託管的開源大語言模型（如Llama 3、Mistral和Gemma）進行文本分類任務。

Mistral Vibe是一款AI代理，專為長期運行的多步驟工作和編程任務設計。本文介紹其功能及在Product Hunt上的討論。

在Mistral AI峯會上，CEO Arthur Mensch表示歐洲必須在兩年內建立足夠的AI基礎設施，否則可能淪為美國AI的“附庸國”。峯會吸引了眾多歐洲企業和政府代表，強調數據主權和開源模型，但歐洲在投資和規模上仍遠落後於美國對手。

本文分享了作者在巴黎Mistral AI Now峯會上的個人見解。Mistral不再只是一家模型公司，而是構建了包含計算、模型、平台和諮詢服務的完整AI堆棧。峯會重點強調了與ASML、BNP Paribas、亞馬遜等企業的合作，而非新模型發佈。Mistral專注於高效、開放和可定製的模型，並支持本地部署，這成為其區別於Anthropic或OpenAI的獨特賣點。小型專用模型是關鍵戰略，例如用於OCR的Document AI、多語言語音的Voxtral和工業機器人的Robostral。主權和本地部署是歐洲企業的差異化優勢，如BNP Paribas和Abanca的案例。此外，奧地利科學院利用Mistral的編程模型Codestral解讀古代紙莎草文獻，展示了AI在人文領域的潛力。總而言之，Mistral的目標並非贏得AGI競賽，而是成為歐洲的全棧AI合作伙伴，提供即時的實際投資回報。

法國初創公司Mistral AI在Digital Realty的巴黎南園區獲得了10兆瓦的計算能力，以擴展其AI基礎設施。

Mistral AI 將其聊天機器人 Le Chat 更名為 Vibe，並將聊天、編程代理和新的工作模式整合在一個品牌下。工作模式可接入 Google Workspace、Outlook、Slack 或 GitHub，獨立處理電子郵件、報告或拉取請求等任務。Pro 套餐價格從 17.99 歐元降至 14.99 歐元，但未明確使用限制。此舉直接對標 OpenAI、Google 和 Anthropic 的代理型產品。

Mistral AI首席執行官Arthur Mensch證實，公司正在探索開發定製芯片以降低基礎設施成本，與OpenAI和Anthropic競爭。這家法國初創公司還宣佈在法國新建推理數據中心，並推出企業智能代理平台Vibe。

本週AI新聞要點：Anthropic公開了此前僅限國防承包商使用的頂級模型Mythos，使五角大樓級AI能力向開發者開放；DeepMind CEO哈薩比斯將AGI時間線提前至2029年；Starlette框架爆出嚴重認證繞過漏洞，影響數百萬AI代理；CrowdStrike等聯合摧毀Glassworm殭屍網絡；法國巴黎銀行與Mistral達成主權AI安全合作；中國限制阿里和深度求索頂尖AI工程師出境；Uber AI預算超支、ClickUp裁員並引入數千AI代理，同時MIT技術評論數據顯示AI暴露崗位失業率更低，奧特曼撤回白領失業預言。