AI・テクノロジー
Mistral Strikes Back | Medium 3.5 × 4-GPU Self-Hosting Unlocked
機械翻訳 / Machine-translated
A
AI Friends
2026年5月2日・50分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"The AI model world belongs only to the US and China" — that assumption may finally be ready to be rewritten.
On April 29, 2026, Paris-based Mistral AI announced its flagship model, "Medium 3.5." It arrives with a striking combination: an SWE-Bench score of 77.6% — enough to breathe down the neck of Claude Sonnet 4.5 — published weights under a modified MIT license, and the ability to run on just 4 GPUs at your own facility. Announced alongside it, Vibe Remote Agents bring "cloud-based autonomous coding" as a standard feature.
This article covers everything in one sweep: Medium 3.5's specs and positioning, what SWE-Bench 77.6% means in practical terms, the disruptive power of open weights that run on 4 GPUs, how Vibe Remote Agents work, the realities of pricing and operations, competitor comparisons, and the impact on the Japanese market.
Medium 3.5 is Mistral's first flagship model that unifies three capabilities in one: instruction following (doing what you tell it to), reasoning (thinking through problems), and coding (writing programs).
Until now, Mistral's lineup centered on purpose-specific models — the compact Mistral 7B, the coding-focused Devstral, and others — but Medium 3.5 covers all of these in a single model. It's like a restaurant that previously needed multiple specialist chefs finally hiring one executive chef capable of handling every cuisine.
The model also features adjustable "depth of thinking" per request, so simple questions get quick answers while hard problems get thorough reasoning. The ability to handle multiple use cases with a single model cuts operational costs and model-selection overhead in one stroke.
Some background on the company. Mistral AI is a French artificial intelligence company founded in Paris in 2023, and it is Europe's foremost LLM (large language model) startup.
CEO Arthur Mensch, formerly of Google's DeepMind, made headlines just nine months after founding the company by matching OpenAI's performance. In March 2026, Mistral raised $830 million to build data centers in Paris and Sweden, cementing its position as a cornerstone of Europe's AI infrastructure.
Widely watched as the "third force" after the US duo of OpenAI and Anthropic and the Chinese players DeepSeek and Qwen, Mistral also has plans to enter Japan before the end of 2026.
The innovations weren't limited to the model itself — new ways to use it were bundled in as well. On the same day, Mistral also released Vibe Remote Agents (a cloud-based asynchronous coding agent) and Le Chat's Work Mode (a business agent that spans multiple tools).
Previous coding AIs have fundamentally run on local PCs, but Vibe flips that concept: "throw it all to the cloud and have it written while you sleep." Rather than cooking at home yourself, it's closer to sending an order slip to a professional kitchen and saying, "Have it ready by tomorrow morning."
Le Chat's Work Mode is a butler-style AI that handles email drafting, research, and multi-tool coordination all within a single conversation. It's available on Pro, Team, and Enterprise plans.
Understanding the benchmark reveals the true weight of that 77.6% score. SWE-Bench Verified is a test compiled from real open-source bug reports on GitHub. The AI is told, "Write code to fix this bug," and it's evaluated on whether the resulting patch actually works. This isn't a theoretical exam — it's a hands-on skills test using real-world problems.
Medium 3.5 recorded 77.6%, surpassing Devstral 2 and Qwen 3.5 397B (a model four times larger), and came within 1 point of Gemini 3.1 Pro Preview (78.8%), a regular in the top tier. It's like a contender who was aiming for second-class suddenly turning up in the finals alongside a first-class titleholder.
Having clearly crossed the threshold of practical usability as a coding AI, it has reached the level where "colleague" is a more apt description than "assistance tool."
Another standout number deserves attention. τ³ (tau-cubed)-Telecom is a benchmark that measures how well an AI agent resolves telecommunications industry support tasks, and Medium 3.5 scored 91.4.
This tests comprehensive agent capability — calling multiple APIs, maintaining conversational context, and completing customer interactions end-to-end. It's an indicator for long-horizon tasks rather than one-shot Q&A. A 91.4 represents the overall ability of a chef who doesn't just memorize recipes but listens to the customer's preferences, retrieves the right ingredients, makes adjustments along the way, and delivers a finished dish.
Beyond coding, the model is fully capable of serving as a business automation agent in enterprise environments.
"How much information can it handle at once" is also a critical metric. Medium 3.5 supports a context length of 256k (256,000) tokens — roughly 200,000 Japanese characters, or more than an entire paperback book, in a single read.
For example, you can feed it an entire codebase (hundreds of files) at once and instruct it to "identify the source of this bug." Rather than a student looking things up in a dictionary while they work, it's more like a student who has internalized every page of the textbook before sitting down to answer questions.
Tasks that previously required "chunking and feeding in pieces" — reviewing lengthy contracts, refactoring massive codebases, referencing the history of long-running projects — can now be handled in a single pass. It exceeds Claude Sonnet 4.5's 200k and GPT-4o's 128k, placing it at the top of the class for long-context processing.
Here is the essence of the license. Medium 3.5 is released under a "Modified MIT License," freely downloadable by anyone from Hugging Face, and usable for both commercial and non-commercial purposes.
However, an exception clause applies to "companies with large-scale revenue," requiring an additional contract in some cases. In essence: small businesses and individual developers get complete freedom; only the very largest enterprises need to go through a separate channel. Think of it as any neighborhood restaurant being free to use the recipe, while only major chains need a separate licensing arrangement.
It follows the same "quasi-open-source" trajectory as LLaMA and DeepSeek, while incorporating compliance considerations befitting a European company — a practical license design.
The hardware requirements are what's generating the most attention. At FP8 (8-bit floating point) precision, Medium 3.5 runs on just 4 datacenter-grade GPUs (NVIDIA H100 80GB × 4, totaling 320GB VRAM).
For a flagship model at the 128B parameter scale, this is remarkably lean — realistic to operate on enterprise on-premises infrastructure within a sensible budget. While GPT-4o and Claude are only available through APIs, Medium 3.5 can be placed on your own servers. This is expected to accelerate adoption in finance, healthcare, and government, where confidential data cannot leave the premises.
Rather than buying out a famous restaurant's chef outright, the analogy is: receive the full recipe and all the kitchen tools, and reproduce it in your own establishment. GPU costs aren't trivial, but over the long term, owning the infrastructure beats paying monthly API fees by a wide margin.
The software side is equally well-equipped. Medium 3.5 is compatible with all the major LLM inference engines — vLLM, SGLang, and Ollama — so engineers can choose based on their preferences.
A separate "EAGLE" model for fast inference is also published, providing the mechanism to multiply token generation speed several times over.
vLLM is a high-performance Python-based server; SGLang excels at structured reasoning; Ollama is a lightweight tool that even runs on a personal Mac. The setup allows everyone from engineers who want to experiment locally to infrastructure teams running production deployments to choose the optimal solution for their environment.
Vibe's innovation in one sentence: "the location of the coding AI has changed." Previous tools like Claude Code, Cursor, and GitHub Copilot fundamentally ran on local PCs. Vibe runs on Mistral's cloud, and the user simply issues instructions. Multiple tasks can execute in parallel simultaneously, and you get notified when they're done.
Instead of standing in the kitchen yourself, picture sending an order slip to a professional kitchen with a full staff of chefs — "these three things, please have them ready by tomorrow morning." That's the mental model shift.
During execution, file diffs (what changed) and tool-call progress are made visible, so there's no anxiety about having handed everything off to an AI with no idea what it's doing. It's accessible via Mistral Vibe CLI (command line) or Le Chat (chat interface) and integrates naturally into existing development workflows.
Whether it's usable in the real world comes down to how it connects with other tools. Vibe ships with out-of-the-box integrations for GitHub (code and pull requests), Linear and Jira (issue tracking), Sentry (error monitoring), and Slack and Teams (chat notifications).
For example: an error surfaces in Sentry, Vibe automatically identifies the offending code, checks out a branch in GitHub, writes the patch, opens a pull request, and reports back in Slack — all in one automated sequence, triggered by a single instruction. Order slip (Jira) → ingredient inventory (GitHub) → cooking (coding) → plating (pull request) → service (Slack notification): the entire pipeline runs on its own.
Engineers only need to communicate what they want built; Vibe handles everything downstream. In that sense, this marks a turning point in the development workflow — one where teams can refocus entirely on requirements and design.
Another compelling feature is "Teleport." It's a mechanism that moves a work session in progress from your local PC to the cloud — with the full history intact.
Say you were writing code in a café before heading out and only got halfway through. Rather than continuing on your desktop when you get home, Vibe on the cloud can keep writing while you're still in transit. Like leaving a dish half-prepped and walking out the door, knowing the rest of the kitchen staff will take it to completion.
The concept turns meetings and commutes into "AI working hours," with the potential to effectively more than double an engineer's productive output. Execution happens inside a sandbox environment (an isolated, safe workspace), so there's no risk of breaking production.
Here is a summary of the pricing. Medium 3.5's API is priced at $1.50 per million input tokens and $7.50 per million output tokens.
Compared to Claude Sonnet 4.5 ($3 input / $15 output) — roughly half the price — and cheaper than GPT-4o ($2.5 input / $10 output) as well. It's a price point that hits hard for cost-conscious users: the taste of a three-Michelin-star restaurant at bistro prices.
For a mid-sized system processing one million tokens of input and output daily, switching from Claude to Mistral could realistically save over ¥300,000 per month. Three plans are available — Pro (for individual developers), Team (for teams), and Enterprise (for large organizations) — so usage can scale to fit the context.
There's also a non-API option. In Mistral's official chat app, Le Chat, Medium 3.5 is available on monthly Pro, Team, and Enterprise plans, and the new Work Mode feature is included.
Accessible directly from a browser or smartphone app — just like ChatGPT — it gives non-technical general users a direct path to Mistral's latest model. The API version is like a direct ingredient delivery; Le Chat is like a finished dish ready to eat — different paths for different use cases.
Japanese input and output are supported, so it's fully usable even for those not comfortable in English. Specific monthly pricing varies by region and plan, so checking Mistral's official website is recommended. The cost feel is comparable to ChatGPT Plus or Claude Pro (around $20/month).
Whether API billing or self-hosting is more economical depends on usage volume. The breakeven point is roughly: API for up to around 100 million tokens per month (small to mid-sized systems); self-hosting on 4 × H100s for over 1 billion tokens per month (large-scale production).
Cloud rental of H100 × 4 runs approximately ¥300,000–¥500,000 per month. For companies handling confidential data, that buys both cost savings and data sovereignty — a double win. It's simply: if you eat out once or twice a month, takeout (API) makes sense; if you eat three meals a day every day, your own kitchen (self-hosting) is cheaper.
The expected split is small and mid-sized businesses on API, large enterprises self-hosting — and Mistral's major differentiator is that it provides the same quality through either route, unlike its competitors.
Starting with Anthropic's flagship. Claude Sonnet 4.5 scores around 77–78% on SWE-Bench and is industry-top rated for reasoning stability and safety measures.
That said, it's priced at $3 input / $15 output — twice Medium 3.5 — and is API-only with no open weights. If Claude is a "top three-Michelin-star French restaurant," Mistral is "the place serving the same food at bistro prices, and you can take the recipe home."
Claude is the safe choice for use cases where confidentiality is paramount (legal, medical, etc.); Mistral has the edge where balancing cost and data sovereignty is the priority. Anthropic's strength in "Constitutional AI" (its proprietary safety methodology) requires custom prompt engineering to replicate in Mistral, making accumulated operational know-how a deciding factor.
The comparison with OpenAI is equally interesting. GPT-4o and GPT-5.5 deliver high overall scores in knowledge, multimodality, and reasoning, and the ChatGPT ecosystem is an overwhelming strength.
The weakness: no open weights, API-only, no self-hosting, and all data must pass through OpenAI's servers — a closed system. OpenAI is "a world-class restaurant chain with locations everywhere"; Mistral is "a European brand with a recipe included so you can reproduce the dish at home."
If you want to outsource AI, OpenAI is the fit. If you want to own AI internally, Mistral is the fit. In Japan, self-hosting demand for Mistral is projected to grow in confidential domains within finance, defense, and manufacturing — enough momentum to put a dent in OpenAI's stronghold.
Google's Gemini also warrants comparison. Gemini 3.1 Pro Preview scores 78.8% on SWE-Bench — just 1.2 points above Medium 3.5's 77.6% — making coding performance essentially a toss-up.
Gemini's strengths are its Google Search and Workspace integrations; its weakness is being a closed system that's difficult to operate in multi-cloud scenarios. Gemini is "the Google-branded chain — earns you points"; Mistral is "the independent with the freedom to source ingredients from any supplier."
For all-in Google shop environments, Gemini; for mixed multi-cloud and on-premises setups, Mistral has the easier fit. Given the negligible score gap, the practical decision point is compatibility with your existing IT infrastructure rather than benchmark numbers.
Mistral's Japan expansion is accelerating. In coverage by the Nikkei Shimbun in May 2026, CEO Mensch stated that the company would "establish a Japan office within the year and capture demand from the manufacturing sector's digital transformation."
The primary targets are major Japanese corporations in automotive, electronics, banking, and similar industries that want to use European-made AI rather than depending on US or Chinese providers. The dynamic is similar to a serious French restaurant finally opening in Japan for discerning diners who've grown a little tired of US and Chinese fast food.
Japan's culturally ingrained preference for confidentiality — keeping internal data from leaving the organization — is a perfect match for Mistral's open-weight strategy. Nikkei Cross Tech has analyzed that the second half of 2026 will see a succession of proof-of-concept (PoC) deployments at major companies including Toyota, MUFG, and NTT.
Japanese language performance is another relevant consideration. Mistral models support Japanese input and output as standard, and both Rakuten and Stability AI Japan have previously released Japanese-specialized versions fine-tuned on earlier Mistral models.
Similar fine-tuned variants are expected to emerge for Medium 3.5, pointing toward significant improvements in natural Japanese. The vision is a French executive chef arriving in Japan and independently developing a Japanese-French fusion cuisine tailored to local palates — regional optimization will proceed apace.
Even the current Medium 3.5 is at a fully practical level for Japanese-language coding instructions and document summarization, but for dialects and specialized industry terminology, waiting for Japanese-specific variants is the cautious approach. The options for Japanese-language LLMs — previously "OpenAI / Anthropic / Google / Chinese players" — now gain "France," ushering in a four-pole era.
The relationship with Japan's domestic LLM initiatives also deserves attention. In April 2026, Japan's Digital Agency released a domestic LLM called "Gennai" as open source, positioning it as the core of the government's AI infrastructure.
Mistral is expected to occupy a "commercial flagship" role while Gennai serves as the "national foundation" — a complementary rather than competitive relationship. Gennai is the public-infrastructure school cafeteria; Mistral is the private-sector neighborhood restaurant. Their roles don't overlap.
Europe and Japan are both moving in the direction of establishing "AI sovereignty independent of the US and China," and they're doing so in sync. The expected division: public bodies and government agencies on Gennai; private-sector enterprises on Mistral. For Japan's AI strategy overall, Mistral's entry is a clear net positive in terms of diversifying options.
Tanaka, CTO of a SaaS startup in Tokyo, switched from Claude Sonnet 4.5 to Mistral Medium 3.5 in May 2026, compressing a monthly API cost of ¥600,000 down to the ¥300,000 range — roughly half.
He also introduced Vibe Remote Agents, auto-assigning internal bug tickets to Vibe so engineers could redirect their focus entirely to specification discussions. "Vibe would investigate 30 Sentry alerts a day by morning, and I'd arrive at the office to find 10 pull requests waiting. Subjectively, the team's productive hours went up by 1.5×," Tanaka reflects.
It's a cost structure like renting a 24-hour professional kitchen for ¥300,000 a month — a complete win on value. The freed-up budget was redirected to new engineer hires, adding another gear to the development organization's growth cycle.
Sato, who handles systems development at a major Tokyo bank, was bound by a constraint: customer account data cannot go anywhere near external APIs. Until recently, the in-house code generation AI was underpowered enough that a sense of resignation had set in.
In May 2026, Medium 3.5 was deployed to an in-house H100 × 4 server, delivering SWE-Bench 77.6% performance internally without a single byte of confidential data leaving the premises. "The modified MIT license cleared legal, the GPU cost estimates cleared the budget review. We can finally start serious AI deployment," says Sato.
A three-Michelin-star chef hired for the company cafeteria — after years of waiting, the finance industry's long-sought option to use cutting-edge AI while maintaining data sovereignty has become reality.
Yamada, a freelance engineer based in Yokohama, began running a quantized version of Medium 3.5 via Ollama on his high-spec home Mac Studio (M3 Ultra, 192GB unified memory) in May 2026.
"Zero API costs, running personal projects at a million tokens a month, and it even works offline," he says. Using Vibe CLI to run overnight batch code reviews on personal projects, he's built a habit of waking up to find refactoring suggestions lined up in GitHub every morning.
It feels like having a Michelin-caliber recipe and a full set of kitchen equipment at home, making serious food entirely on your own. For freelancers and independent developers, the era of playing with the latest AI without worrying about API fees has finally arrived.
A. The Le Chat monthly plan is the most accessible, and for engineers, Vibe CLI is the natural first choice.
Non-engineers can simply sign up for Le Chat in a browser and use it just like ChatGPT. Engineers can install Vibe CLI locally, connect it to the API, and fold it into their coding workflow — the fastest path. Self-hosting requires GPU costs of several hundred thousand yen per month, so it's only practical for very high-volume usage or situations involving confidential data.
Think of it as three tiers: eating out (Le Chat) → delivery (API) → serious home cooking (self-hosting). Start with Le Chat to get a feel for it, then move to API or self-hosting as your needs grow.
A. For bug fixes and refactoring on mid-sized web applications, it's ready to pull its weight immediately. For large-scale new development from scratch, it still plays a supporting role — that's the honest answer.
Reading an existing codebase, identifying bugs, and writing patches that pass tests: this is already at or above human-quality output. On the other hand, new projects starting from business requirements still can't be fully handed off to AI — there are phases where a human architect's judgment is essential. The recipe-following is flawless; creative cooking still requires human collaboration — keeping that boundary in mind is key.
Practical tips: "break tasks into small pieces before handing them to AI" and "always have a human do the final review." These two habits alone are producing reported cases of development velocity increasing 2–3× across teams.
A. The key point: you can download the model's "internals" and run them on your own servers — and modifications are permitted too.
The weights (trained parameters) are downloadable from the official Hugging Face page, and fine-tuning (additional training) for your organization's specific needs is allowed. This is the biggest difference from API-only models like Claude and GPT, and the ability to use AI without data ever leaving your environment is a decisive advantage.
It's like a restaurant whose recipe was previously secret finally making it public — you're now free to reproduce it at home and adapt it however you like. Two caveats: "large-revenue enterprises require a separate contract," and "GPU costs for running it are the user's responsibility." For small businesses and individuals, it's completely free to use.
A. Individual purchase is a stretch — cloud rental is the realistic option.
A single NVIDIA H100 80GB runs about ¥4.5 million; four of them tops ¥18 million. They also require datacenter-grade power and cooling, making home installation essentially impossible. On cloud, an H100 × 4 instance on AWS, Azure, or GCP runs roughly ¥3,000–¥5,000 per hour; 100 hours a month would cost ¥300,000–¥500,000.
For individual users, another option is running a quantized version (a lighter model with reduced precision) on a Mac Studio or RTX 4090 × 2 setup — via Ollama, you can get practical speeds on a PC costing a few hundred thousand yen. The three choices are: buy a professional kitchen, rent one, or make do with a home stove. Choose based on your use case and budget.
"The AI race is settled between the US and China" — that assumption is now ready to be overwritten.
Medium 3.5, announced by Mistral on April 29, 2026, is Europe's AI counteroffensive in concrete form: SWE-Bench 77.6% coding performance delivered as open weights under a modified MIT license. Self-hostable on 4 GPUs, API-priced at roughly half of Claude Sonnet 4.5, and now capable of cloud-based asynchronous coding through Vibe Remote Agents. With Mistral's Japan entry before the end of 2026, this is fast becoming the go-to option for every organization that needs to balance confidentiality and cost.
Three things you can do starting today: ① Experience the capability firsthand at Le Chat (chat.mistral.ai); ② Install Vibe CLI and test coding integration; ③ Audit your organization's confidentiality requirements, API costs, and GPU assets, then build a comparison table against Claude and GPT. Europe's AI is serious — and it is already beginning to reshape the development landscape in real and measurable ways.
This article is a cross-post from AI Friends.
