Mistral Large
Mistral AI’s launch of its new flagship “Mistral Large” model, offered as a paid API and on Azure, is seen as a strategic move into the top tier of commercial LLMs alongside GPT‑4 and Gemini. Commenters focus on its slightly lower pricing and near‑state‑of‑the‑art benchmarks, but many are disappointed that, unlike earlier Mistral releases, the new model weights are closed and the company’s branding has shifted away from an explicit commitment to open weights. This shift fuels broader concerns that “open source” in AI is often a temporary marketing strategy until vendors can lock down their strongest models, with implications for long‑term competition, research, and user control.
Open vs Closed and “Open Source” Positioning
- Many see Mistral’s new closed models (and website copy change from “Open-Weight models” to “Frontier AI in your hands”) as a pivot away from its earlier open-weights branding, likening it to OpenAI’s trajectory.
- Some argue “open source” does not really apply to opaque weight blobs; others counter that weights are the “preferred form for modification” in ML and thus reasonably count.
- Several comments frame open releases as a marketing or “freemium” stage until a model is competitive enough to sell.
- A minority defends Mistral: companies must fund expensive training, and giving away strong open models may mainly benefit “copy-cat” competitors.
Pricing, Azure Integration, and Enterprise Angle
- Mistral Large pricing: $8 / 1M input tokens, $24 / 1M output tokens—slightly cheaper than GPT‑4‑Turbo ($10 / $30).
- On Azure, Mistral is marginally cheaper than GPT‑4‑Turbo; classic GPT‑4 is much more expensive.
- Some wonder why anyone would use Mistral over GPT‑4 when it benchmarks slightly worse yet is only ~20% cheaper.
- Others note Azure distribution is key for enterprise: existing contracts, compliance (e.g., FedRAMP), and indemnification make “just another Azure service” far easier to buy than a startup’s own hosting.
Capabilities, Benchmarks, and Real‑world Tests
- Mistral’s own charts show it slightly below GPT‑4 and (per commenters) below Gemini 1.5 / Ultra on some benchmarks; choice of Gemini Pro 1.0 as comparator is criticized as outdated.
- Users report mixed hands-on results: some find Le Chat better than GPT‑4 on specific coding tasks; others still see GPT‑4 as clearly superior in reasoning, code quality, and understanding intent.
- Mixtral 8x7B continues to be praised as an exceptionally useful open model, especially for local use, with some preferring it over Google search for technical queries.
API, Naming, and Product Line Changes
- Endpoints were renamed and versioned (e.g.,
open-mistral-7b,open-mixtral-8x7b,mistral-small-2402,mistral-large-2402), with older aliases scheduled for deprecation. - New features include function calling and JSON mode for Mistral Small and Large, a new Le Chat interface, and enterprise account management.
- Some interpret the split between “open-*” and proprietary endpoints as a signal that future high-end models will stay closed.
Economic and Ecosystem Concerns
- Debate over whether releasing weights would meaningfully hurt revenue: some think infra and sales moats would preserve OpenAI/Mistral’s business; others point to how quickly third parties host open models cheaper and faster.
- Fears expressed about concentration of top models under a few big players (notably Microsoft‑aligned), potential antitrust/“enshittification,” and the lack of robust community‑funded open training efforts.