Reflection AI vs Mistral vs DeepSeek — what's the difference?

Three regional bets on open-source frontier capability. Reflection AI (US, $25B valuation, founded by ex-DeepMind researchers, just signed $6.3B SpaceX compute deal on June 22, 2026) is the American open-source frontier program — no public model yet but the funding and compute commitment are real. Mistral (France, founded 2023) is the European sovereign-AI play with Mistral Large 2 currently shipping and EU regulatory tailwinds. DeepSeek (China) shipped V4 Pro earlier in 2026 with frontier-competitive benchmarks and is the most credible 'use today' open-weights option. They serve different customers: Reflection courts US government and regulated enterprise, Mistral courts EU sovereignty buyers, DeepSeek serves global cost-sensitive and AI-research users.

Which open-source frontier model can I actually use today?

DeepSeek V4 Pro (China) is the most capable open-weights frontier model you can download and run today. Mistral Large 2 (France) is the most polished — best documentation, broadest hosted-inference availability, cleanest licensing for commercial use, but slightly less frontier-edge on raw capability. OpenPangu-2 (Huawei) and Kimi K2-7 Code (Moonshot) and GLM 5.2 (Zhipu) are other strong Chinese open-weights options for specific use cases (coding-heavy, long-horizon). Reflection AI has no released model yet. If you want US-origin frontier open-weights right now, the honest answer is: nothing exists yet — wait for Reflection's first ship, or use Mistral as the closest non-Chinese alternative.

Is Reflection AI's $6.3B SpaceX deal enough to catch Anthropic / OpenAI?

Compute is necessary but not sufficient. The $6.3B SpaceX commitment is real: GB300 chip access at Colossus 2 is roughly equivalent to a meaningful fraction of what Anthropic ($1.25B/mo, $52.5B total) and Google ($920M/mo, $38.6B total) have. The smaller absolute size means slower scaling, not impossible scaling. What matters more: (1) talent — Reflection has ex-DeepMind founders, but Anthropic just hired John Jumper and the talent war is fierce, (2) execution — Reflection has shipped nothing yet, while OpenAI shipped GPT-5.5-Cyber and Patch the Planet today, (3) revenue path — Reflection needs DOE / DoD / banks to materialize as paying customers to fund the open-source give-away. Realistic timeline: a credible Reflection frontier model arrives late 2026 or early 2027.

Where does Llama / Meta fit in?

Increasingly nowhere. Meta announced earlier in June 2026 that it is closing the Llama / Muse / Spark open-source program and pivoting away from frontier open-source. Llama 5 is effectively the end of the line. The reason: Meta concluded that open-weights frontier doesn't generate enough product revenue to justify the training spend at frontier scale. This vacated the 'American open-source frontier' slot that Reflection is now stepping into — and is the most direct strategic context for the Reflection / SpaceX deal. If Meta had stayed in the game, Reflection's $25B valuation would look much more crowded. With Llama 5 closing, Reflection is essentially uncontested as the US open-source frontier bet.

Quick Answer

Reflection AI vs Mistral vs DeepSeek — Open-Source Frontier Race (Jun 2026)

Published: June 23, 2026

Reflection AI vs Mistral vs DeepSeek: The 2026 Open-Source Frontier Race

Three regional bets on open-source frontier AI capability, each with different funding, different shipping status, and different target customers. The June 22, 2026 SpaceX deal puts Reflection AI on the board for the first time. Here’s how the three compare, what’s actually usable today, and where each is going.

Last verified: June 23, 2026.

TL;DR

Lab	Region	Valuation	Latest model	Compute commit	Best customer
Reflection AI	US	$25B (Mar 2026)	None public yet	$6.3B SpaceX (Jun 22)	US gov, regulated enterprise
Mistral	France / EU	Significant; private	Mistral Large 2	EU partnerships	EU sovereignty buyers
DeepSeek	China	Private	V4 Pro	Chinese domestic stack	Global cost-sensitive + research

Reflection AI — US, no model yet

The new entrant. Founded by two former Google DeepMind researchers, valued at $25 billion in March 2026 reporting, backed by Nvidia, with JPMorgan Chase reportedly considering investment via its Security and Resiliency portfolio. On June 22, 2026, Reflection signed a $6.3 billion compute deal with SpaceX — $150M/month for GB300 capacity at Colossus 2 near Memphis from July 2026 through 2029.

Strategy: Be the American open-source frontier lab. US governments, banks, and regulated enterprises want frontier capability with open weights — they won’t deeply commit to closed US labs (sovereignty) or Chinese open-weights (security). Reflection is the third option.

Government ties: DOE Genesis Mission participation confirmed; other national security clients likely.

Risk: No public model yet. The thesis depends on Reflection shipping a credible frontier model in late 2026 or early 2027. The compute commitment is the bet, not the proof.

Mistral — EU, mature open-weights

The European leader. Founded 2023 in France, Mistral has shipped a steady cadence of strong open-weights models culminating in Mistral Large 2. Mistral Large 2 isn’t the absolute frontier on every benchmark — that title rotates among GPT-5.5, Claude Fable 5, Gemini 3.5 Pro, and DeepSeek V4 Pro depending on benchmark — but it’s the best polished open-weights product for European enterprise.

Strategy: EU sovereignty positioning. Mistral is the EU’s answer to American AI dominance, with regulatory and procurement tailwinds. Recent EU AI Act compliance momentum and the August 2026 deadline are real customer drivers.

Strengths: Best documentation in the open-weights world, cleanest commercial licensing, broadest hosted-inference availability, strong EU enterprise relationships.

Weakness: Capability gap vs frontier closed models is real and widening. Mistral has to keep shipping fast to stay in the conversation.

DeepSeek — China, “use today” frontier open-weights

The most capable open-weights frontier model you can download today. DeepSeek V4 Pro (shipped earlier in 2026) competes with frontier-tier closed models on a range of benchmarks — particularly coding, math, and reasoning. The DeepSeek family’s training efficiency has consistently surprised Western labs.

Strategy: Global cost-sensitive frontier capability + research community goodwill. DeepSeek hasn’t been positioning as a sovereign-AI play; it’s been positioning as “frontier capability with open weights, period.”

Strengths: Best raw capability among open-weights options as of June 2026. Strong with researchers and developers. Aggressive scaling roadmap.

Weakness: Trust concerns from US enterprise and government use. Some procurement organizations have explicit “no Chinese AI weights” rules. This is the gap Reflection is trying to fill.

Other Chinese open-weights — context

DeepSeek isn’t alone in the Chinese open-weights cohort:

OpenPangu-2 (Huawei) — frontier-competitive, tightly integrated with the Chinese domestic compute stack
Kimi K2-7 Code (Moonshot) — coding-strong, popular with developers
GLM 5.2 (Zhipu) — long-horizon coding strength

This cohort collectively raises the bar for what “open-weights frontier” means. Reflection AI’s first model has to land in this neighborhood to be credible.

Meta / Llama — the vacated slot

Meta announced in June 2026 that it is closing the Llama / Muse / Spark open-source program. Llama 5 is the end of the line. The reasoning: open-weights frontier doesn’t generate enough product revenue to justify training spend at frontier scale.

This is the strategic context that makes Reflection’s bet make sense. With Meta out, the “US open-source frontier” slot is genuinely vacant. Whoever fills it captures a real customer segment (US gov, regulated enterprise, allied governments) that won’t accept closed US labs or Chinese open-weights.

Which should you bet on?

If you’re a developer using AI today: DeepSeek V4 Pro or Kimi K2-7 Code (coding) or GLM 5.2 (long-horizon). They’re the highest-capability open-weights options shipping.

If you’re an EU enterprise: Mistral Large 2, plus watch for Mistral’s next major release. The sovereignty story matters more than the absolute capability gap.

If you’re a US government, defense, or regulated enterprise: Watch Reflection. If they ship a credible model in late 2026 / early 2027, it likely becomes your default — the procurement and policy fit is too clean to ignore. Until then, you’re stuck with closed US labs.

If you’re investing: Reflection is the highest-asymmetry bet (early, large funding, vacant strategic slot, but no model). Mistral is the most de-risked option with clear EU revenue. DeepSeek is private and harder to access; OpenPangu / Kimi / GLM are similar.

Sources

CNBC, Yahoo Finance, MLQ News coverage of Reflection / SpaceX deal, June 22, 2026
WSJ reporting on Reflection $25B valuation (March 2026)
DeepSeek V4 Pro public release coverage
Mistral Large 2 documentation
andrew.ooo prior coverage on Llama 5 closing and Chinese open-weights cohort

Verified June 23, 2026.