# BOTS FIRED

WEDNESDAY, OCTOBER 7, 2026

**Mistral previews Large 4 as Europe’s open-weight contender.**

This batch is about control of the AI stack: models, local retrieval, simulation data, power, biology data and workforce graphs. Most capability claims remain vendor-reported, while the money and infrastructure commitments are more concrete. The recurring operational question is who gets the model, data and electricity first.

# 🤖 FRONTIER MODELS

## Mistral previews Large 4 before releasing its weights

**Mistral has opened a public API preview for its 1 trillion-parameter Mistral Large 4 model.**

The company says the multimodal mixture-of-experts model has 49 billion active parameters and is its largest model yet. It is available through Mistral Studio during the preview, while the weights are not yet released with release pending. Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities before that release.

In its [public model preview announcement](https://mistral.ai/news/mistral-large-4/), Mistral reports benchmark and evaluator results across coding, cybersecurity, finance, law and visual grounding. Those comparisons include internal tests and third-party evaluations, so they are useful claims rather than settled procurement evidence. The business pitch is sovereignty: private-cloud or on-premise deployment under customer policies. The release plan does not establish that deployment will be simple, cheap or safe. Customers still have to validate it.

## Reflection puts Beam’s 501B open-weight model into early access

**Reflection has introduced Beam, a 501 billion-parameter sparse model with 23 billion active parameters.**

Beam is undergoing final red-teaming and evaluations, and Reflection is accepting early-access signups. The company says it will release the weights, model card, technical report and developer artifacts after final evaluations. That makes Beam a pending open-weight release, not a generally available set of weights.

In its [Beam model announcement](https://reflection.ai/blog/introducing-beam), Reflection reports results for coding, agentic work, reasoning and tool use. It says Beam uses three to four times less inference compute than some comparison models on selected reasoning tasks, based on approximate FLOP estimates that exclude prompt prefill, attention operations and serving overhead. Reflection also describes a reinforcement-learning run using 10,500 NVIDIA GB300 GPUs over four weeks. The business implication is an efficiency pitch for enterprise agents, subject to final evaluations and independent reproduction. For now, the model is a well-documented RSVP.

# 🦿 ROBOTICS & PHYSICAL AI

## Reka previews Rho-1 as a single model for video and actions

**Reka has released Rho-1 as a research preview of a 19 billion-parameter omni-reasoning model.**

The company says one network handles text, images, video and robotic actions in a shared context, rather than passing work between separate specialist models. Its examples show a session that generates a lighthouse image, marks it, animates it, changes the weather and describes the edit. Reka also presents robotic-control demonstrations and a model architecture that treats actions as continuous tokens.

The [research post](https://reka.ai/labs/research/rho-1-collapsing-the-multimodal-stack) includes vendor demos and internal timing claims, not production deployment evidence. Reka acknowledges limitations: long rollouts can drift structurally, video grounding is not yet reliable across time, targeted edits are brittle, and native video is capped at 672 by 384 resolution. The implication is that unified multimodal state could reduce handoffs in experimental systems. It does not establish reliable robot operation outside the demonstrated settings. The robot still gets a research-preview badge, not a factory shift.

# 🔓 OPEN SOURCE

## Google releases EmbeddingGemma 2 for local multimodal retrieval

**Google has released EmbeddingGemma 2, a 740 million-parameter multimodal embedding model under Apache 2.0.**

The model maps text, code, images, audio and video into one embedding space, according to Google. It can use a 270 million-parameter text-only configuration, with optional vision and audio encoders. Google says quantized text-only weights need about 191MB of active RAM on a Pixel 11 Pro, while the full multimodal configuration needs about 567MB.

The [release post](https://deepmind.google/blog/embeddinggemma-2-an-open-lightweight-multimodal-embedding-model/) positions it for on-device search, retrieval and routing, including offline media search and local RAG pipelines. Google reports leading sub-1B results on several embedding benchmarks, including a code-score improvement over the first EmbeddingGemma. Those are vendor-reported benchmark results. The practical implication is that developers can test cross-modal search without sending every file to a remote embedding API. Their vector-store bill may still find a way to attend the meeting.

# 🧬 SCIENCE & HEALTH

## OpenAI shares mathematics results from an unreleased frontier model

**OpenAI has published mathematical results generated by an internal frontier model that it has not released.**

The company placed results in a GitHub repository with revision and citation protocols after consulting an independent advisory group at the Institute for Advanced Study. Many proofs include Lean formalizations for computer checking, and OpenAI says more formalizations will follow. The repository also contains ten summaries of model reasoning, estimated compute use and counts of attempted problems.

According to [OpenAI’s announcement](https://openai.com/index/sharing-ai-progress-in-mathematics), average result compute was roughly equivalent to three hours of ChatGPT Pro thinking. OpenAI says exposition, citations and presentation remain areas for improvement. The release makes selected output and checking materials available for scrutiny, but it does not make the underlying model available and does not establish independent validation of every result. Research groups can inspect proofs and processes while access to the model itself remains unresolved. Mathematics gets the audit trail; the model stays backstage.

## Biohub coordinates $1.8bn for AI-ready biology data

**Biohub describes a $1.8 billion combined effort to supply data for AI models that predict cell behavior.**

The stated total includes funding, existing pledges, datasets, computing and new measurement technology, rather than $1.8 billion of newly raised cash. Meta, Google DeepMind and Isomorphic Labs are contributing $300 million between them, while Biohub’s own earlier $500 million pledge is part of the wider effort. The Department of Energy plans more than $500 million over five years, and the NIH contributes repositories built with earlier federal funding.

The [AI biology data commitment](https://thenextweb.com/news/biohub-1-8bn-ai-biology-data-meta-google-deepmind) says the resulting datasets will become public resources. Commercial funders receive one year of exclusive access to data covered by that arrangement; government-funded work has no such restriction. Biohub plans a first dataset in about a year. The implication is a larger shared data pipeline for biological models, with a paid head start for some participants. Open data arrives, just not simultaneously for everyone.

# ⚡ ENERGY & INFRA

## Google signs nuclear upgrade agreement for data-center power

**Google has signed a 20-year agreement with Constellation to support upgrades at six US nuclear plant sites.**

The companies expect equipment upgrades across 11 reactors in Illinois, Pennsylvania and New Jersey to add 890 megawatts of nuclear capacity over five years. Constellation says the arrangement represents more than $4.3 billion of new investment. The sites connect to the PJM grid, which serves a major concentration of US data centers.

The [nuclear power purchase agreements](https://www.theverge.com/science/1006082/google-nuclear-energy-power-purchase-agreement-constellation) describes planned capacity, not power already delivered. Google also made a separate 15-year, 2,700MW energy-supply agreement with Constellation for existing PJM plants regardless of fuel type. The story places the deals against rising regional demand and concerns about costs for other electricity customers. The business implication is that hyperscale computing is becoming a long-term utility procurement problem as much as a chip problem. The atoms are scheduled; the extra megawatts are not yet on the meter.

# 💸 DEALS & FUNDING

## Vinci raises $250m for AI chip and hardware simulation

**Vinci has raised $250 million at a $1.5 billion valuation for physics simulation software.**

Advent, Temasek and Xora Innovation led the round, with earlier investors Eclipse and Khosla Ventures participating again, the company said. Vinci simulates physical behavior before hardware is built and initially focused on heat in increasingly complex AI chips. It now aims to simulate entire systems and later add areas such as vibration testing and electromagnetics.

The [chip simulation funding announcement](https://thenextweb.com/news/chip-simulation-startup-vinci-raises-250m-at-a-1-5bn-valuation) says Vinci intends to spend the capital on compute, hiring and more simulation products. Chief executive Hardik Kabaria said the company wants to increase pilot deployments with customers from two to 20. Vinci says its model runs simulations up to 1,000 times faster than conventional tools without losing accuracy, but that claim has not been independently verified. The implication is that investors see room for a new simulation layer alongside established engineering-software vendors. Faster simulation is attractive; validation still pays rent.

## SAP agrees to acquire TechWolf for workforce skills mapping

**SAP has agreed to buy Belgian workforce-data company TechWolf, subject to regulatory approval.**

The terms were not disclosed, and closing remains pending. Until closing, SAP and TechWolf will continue to operate separately. TechWolf builds an updated model of work tasks, employee skills and the external labor market from HR and business systems already used by a company.

The [workforce AI acquisition agreement](https://thenextweb.com/news/sap-acquires-techwolf-ai-workforce) says SAP plans to place the technology in SuccessFactors and its Joule assistant. Under SAP’s stated plans, TechWolf will remain an independent unit under its chief executive in Ghent, and its platform will remain open to non-SAP customers. SAP describes TechWolf’s context graph as grounding for agent queries about skills planning and talent management. The implication is that workforce data becomes a more direct input to enterprise AI products, if the transaction closes and integration follows the plan. For now, it is an agreement, not an acquisition completed in the ledger.

# KEY TAKEAWAYS

**The API preview is public** - the model weights remain pending release.

**Beam remains in final evaluation** - its weights and technical artifacts are still pending.

**Rho-1 is a research preview** - its demonstrations are not proof of production robotics reliability.

**EmbeddingGemma 2 is Apache 2.0 licensed** - local multimodal retrieval can use open weights.

**The underlying model is unreleased** - the public material is results, protocols and some formalizations.

**The $1.8bn total includes existing pledges, data and compute** - it is not all new cash.

**The 890MW increase is planned over five years** - it is not current nuclear output.

**Vinci raised $250m at a $1.5bn valuation** - its 1,000-times speed claim remains unverified.

**The acquisition is agreed but not closed** - regulatory approval remains required.

BotsFired · Bots Keep Firing 🔥