Pixtral Large 2502
Pixtral Large 2502 (mistralai/pixtral-large-2502) is the February-2025 checkpoint of Mistral AI's 124B-parameter open-weight multimodal model — a dense decoder-only transformer paired with a Pixtral ViT vision encoder, 128K context, text output only. It is a documentation/date-bump revision of Pixtral Large 24.11 with no published, variant-specific numeric benchmark table; AWS Bedrock serves it as "Pixtral Large (25.02)" with qualitative best-in-class claims on MathVista, DocVQA and VQAv2 but no scores.
Architectural Profile & Capabilities
Pixtral Large 2502 is Mistral AI's February 2025 revision of the Pixtral Large line (API-era identifier pixtral-large-2502 / pixtral-large-latest; Bedrock model name "Pixtral Large (25.02)"; OpenRouter slug mistralai/pixtral-large-2502). Mistral's own documentation still anchors the public record to the November 18, 2024 release, Pixtral Large 24.11 (pixtral-large-2411), which is now flagged deprecated in Mistral's model catalogue; there is no Mistral changelog entry enumerating feature deltas for a distinct 25.02 checkpoint. AWS, by contrast, does list a separately named 25.02 SKU — it entered Bedrock via custom model import on 9 April 2025 and became a fully managed serverless offering on 17–18 April 2025 — so enterprises most often encounter this exact variant through Bedrock Converse/InvokeModel rather than the first-party Mistral API. Architecturally the model is a 124B-parameter, dense (non-MoE, non-reasoning) autoregressive decoder transformer: the Mistral Large 2-era 123B text backbone plus a ~1B Pixtral ViT vision encoder that accepts variable-resolution, non-square images without forced letterboxing. Both Mistral and AWS document 124B total parameters, 128K context, text+image input and text-only output. The 25.02 revision preserves the same headline specification set as 24.11 (128K context, 124B parameters, image input), which is why the community consensus — and the practical enterprise reading — is that this is a maintenance/date-bump checkpoint focused on serving stability, tokenizer/chat-template consistency and provider packaging rather than a capability step change. No public MoE routing, attention-variant (MLA/GQA) breakdown, or per-image token formula has been published for this exact checkpoint, so any architecture claim beyond "124B dense multimodal transformer with a Pixtral ViT encoder" should be treated as unsourced. Benchmark posture is the critical integrity issue for this dossier. Every numeric figure in circulation (MathVista 69.4, DocVQA ~94.8, ChartQA ~88.1, VQAv2 ~80.0, MM-MT-Bench wins vs Claude 3.5 Sonnet / Gemini 1.5 Pro / GPT-4o) descends from the November 2024 Pixtral Large 24.11 announcement and model card (Pixtral-Large-Instruct-2411) — not from a 25.02 checkpoint evaluation. The Bedrock documentation for Pixtral Large 25.02 asserts "best-in-class performance on key benchmarks including MathVista, DocVQA, and VQAv2" with no numbers attached. Aggregators split accordingly: OneAtlas explicitly states "No published benchmark scores for this model yet" and SuperIndex records "No usable open benchmark results found." Per exact-variant discipline, no benchmark scores are reported in this dossier; the benchmarks array is intentionally empty. Operationally the model is a legacy-tier vision-document specialist. On OpenRouter it runs at roughly 49–72 median output tokens/second depending on the underlying provider (typical cluster ~68–72 tps), and Mistral's own model catalogue rates Pixtral Large 1/5 on speed — deliberate, throughput-constrained serving on a 124B dense backbone. Output ceilings are constrained: Bedrock and aggregators list ~8K max output tokens against the 128K window. Lifecycle risk is material: Mistral marks Pixtral Large deprecated and retired the first-party API version on 31 May 2026, leaving Bedrock and third-party aggregators as the surviving paths. Treat 25.02 as a stable, well-understood OCR/document-QA engine for cost-sensitive batch extraction, not as a frontier reasoning or agentic-coding model — in October 2026 it is two-to-three capability generations behind GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro and DeepSeek V4 Pro on reasoning, coding and agentic work.
124B total parameters, documented by both Mistral AI and Amazon Bedrock; no Mixture-of-Experts routing and no publicly specified expert count, so it is treated as a dense 124B model rather than a sparse MoE. Native multimodal input: text plus images of variable resolution and aspect ratio handled by the Pixtral ViT encoder without forced square resizing or letterboxing (the original Pixtral ViT innovation), enabling documents, charts, tables and natural images. 128,000-token context window with a text-only output channel; ~8K max output tokens on Bedrock/aggregator endpoints. No test-time reasoning/extended-thinking token mode — single-pass generation only. Attention variant (MLA vs GQA), per-image token expansion formula and exact layer/head configuration have not been published for the 25.02 checkpoint and are not asserted here. Function calling and structured outputs (JSON schema) are supported by the Mistral API surface; 25.02-specific tool-call constraints (parallel calls, max tool count, vision+tools coexistence) are undocumented.
Recommended Workloads & Primary Use Cases
- •CONTEXT vs OUTPUT ASYMMETRY: the window is 128K input but output is capped near 8K tokens on Bedrock and most aggregators — long-document summarization or transcript generation must be chunked, and very large multi-page PDFs consume the budget fast because each page image expands into many vision tokens.
- •NO PUBLISHED 25.02 BENCHMARKS: all circulating numbers (MathVista 69.4, DocVQA ~94.8, ChartQA ~88.1, VQAv2 ~80.0) originate from the 24.11 checkpoint and cannot be assumed identical for 25.02; run your own evaluation harness before committing to accuracy SLAs.
- •LIFECYCLE/DEPRECATION RISK: Mistral marks Pixtral Large deprecated and retired the first-party API version on 31 May 2026; only Bedrock ("Pixtral Large (25.02)") and third-party aggregators survive. Treat any production dependency as frozen and plan a migration path to a newer multimodal model.
- •THROUGHPUT AND TTFT: a 124B dense backbone yields roughly 50–75 tps with no reasoning-token mode; time-to-first-token degrades sharply with image count and page resolution, so interactive UIs should pre-resize images and stream partial output.
- •MODALITY METADATA MISMATCH AND TOOL LIMITS: several catalogues (including the OpenRouter modality field) label this model text-only even though it accepts images — validate image payload routing in each provider path. Function calling exists but parallel tool calls, vision-plus-tools combinations and strict-schema limits are undocumented for 25.02; do not assume agentic reliability comparable to modern frontier models.
- •WEAK ON MODERN TASKS: no test-time reasoning means degraded multi-step math, code generation and agentic/tool-orchestration accuracy versus 2026 frontier models; keep it scoped to perception-heavy, single-pass extraction work.
Calls to Pixtral Large 2502 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Mistral AI.
More Models from Mistral AI
Compare other engines available in this laboratory.
Mistral Large 4 ("ML4", internally nicknamed "le Chonk") is Mistral AI's 1.05T-parameter natively multimodal sparse MoE flagship, released as a public preview on October 6, 2026, with 52B active parameters, a 1.6B vision encoder, a 1M-token context window, up to 256K output tokens, and a rare strength in cybersecurity/agentic-automation evaluations.
Mistral Medium 3.5 is a dense 128B instruction-following model from Mistral AI.
Ministral-14b-2512 is a high-performance foundation model engineered by Mistral AI, accessible with Zero Data Retention on ARMES.