MiniMax-M2.5

TL;DR

MiniMax M2.5 is an open-weights agentic LLM (230B-total / 10B-active MoE; architecture unchanged from M2) released Feb 12 2026, positioned as a low-cost frontier alternative to Claude Opus 4.5/4.6 on coding, search, tool-calling, and office tasks. Trained primarily by scaling agentic RL across “hundreds of thousands” of real-world environments using MiniMax’s in-house Forge framework, M2.5 reports 80.2 SWE-Bench Verified, 51.3 Multi-SWE-Bench, 76.3 BrowseComp, and 76.8 BFCL multi-turn, while running at 100 tok/s for the Lightning variant — claimed 1/10 to 1/20 the output-token cost of Opus / Gemini 3 Pro / GPT-5. The Forge framework, the staged parallelism-vs-correctness reward design, and the company-internal data flywheel (“30% of MiniMax’s daily tasks autonomously done by M2.5”) are the technically novel parts; a separate technical blog post on RL scaling is promised.

Key claims

M2.5 reports 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench, and 76.3% on BrowseComp (with context management), described as SOTA in coding, agentic tool use and search, office work, and other economically valuable tasks [§Introduction].
End-to-end SWE-Bench Verified runtime is 22.8 minutes per task (down from 31.3 min on M2.1, a 37% reduction), matching Claude Opus 4.6’s 22.9 min at ~10% of Opus 4.6’s per-task cost [§Efficiency].
Two variants ship with identical capability but different throughput: M2.5-Lightning at 100 tok/s ( $0.3/M input,$ 2.4/M output) and M2.5 at 50 tok/s (half that price); $1/hour continuous at 100 tok/s,$ 0.30/hour at 50 tok/s — framed as “four M2.5 instances running continuously for an entire year for $10,000” [§Cost].
Cross-scaffold generalization is foregrounded: on SWE-Bench Verified with Droid scaffolding M2.5 scores 79.7 vs Opus 4.6’s 78.9; on OpenCode 76.1 vs 75.9 — reported as out-of-distribution-harness evaluation [§Coding].
M2.5’s search behavior is “no longer just getting the answer right, but also reasoning towards results in more efficient paths” — ~20% fewer rounds than M2.1 across BrowseComp, Wide Search, and the internal RISE benchmark [§Search and Tool calling].
An internal Cowork Agent / GDPval-MM evaluation framework pairwise-compares full trajectories (quality of deliverable + professionalism + token cost); M2.5 reports a 59.0% average win rate against mainstream models on office-work scenarios [§Office work].
RL scaling is identified as the primary driver: “hundreds of thousands” of real-world environments built from the company’s own task workspaces, trained under the Forge framework with ~40× training speedup via asynchronous scheduling + tree-structured sample-merging [§RL Scaling, §Forge].
Algorithm-side: continues the CISPO algorithm (from MiniMax’s earlier work) for MoE training stability; adds a process reward mechanism for end-to-end credit assignment in long agent rollouts; and bakes task-completion time into the reward to trade intelligence against response speed [§Agentic RL Algorithm and Reward Design].
A self-deployment data flywheel is reported as an explicit input: 30% of MiniMax’s daily company-wide tasks are autonomously completed by M2.5 across R&D, product, sales, HR, and finance, with M2.5-generated code accounting for 80% of newly committed code [§MiniMax Agent].
Office Skills (typed loaders for Word, PowerPoint, Excel) are integrated into MiniMax Agent’s MAX mode and can be composed with domain SOPs to create reusable Experts; 10,000+ Experts have been built by users [§MiniMax Agent].
Auxiliary benchmark numbers (Appendix table): AIME25 86.3, GPQA-D 85.2, HLE-without-tools 19.4, SciCode 44.4, IFBench 70.0, AA-LCR 69.5 — competitive on AIME25/GPQA but trailing Opus 4.6 / Gemini 3 Pro / GPT-5.2 on HLE and SciCode [Appendix].

Method

Architecture is inherited from M2 unchanged: a sparse MoE with 230B total / 10B active parameters, multi-head causal self-attention + RoPE + QK RMSNorm, top-k expert routing (the architectural specifics are surfaced in NVIDIA’s subsequent M2.7 blog and the M2 series description). The interesting axis for M2.5 is post-training, not architecture.

The post-training recipe centers on agentic RL at scale, executed via the in-house Forge framework. Forge is described as an “agent-native” RL framework: an intermediary layer that decouples the training–inference engine from the agent scaffold, allowing arbitrary scaffolds and tools to be plugged in. Throughput is engineered along two axes: (a) asynchronous scheduling to balance off-policyness against utilization, and (b) a tree-structured merging strategy for training samples — together yielding a claimed ~40× training speedup. The RL objective continues to use CISPO (introduced by MiniMax in early 2025) for MoE training stability, with two additions: a process-reward mechanism for credit assignment in long agent rollouts, and a task-completion-time signal that trades capability against latency.

Training environments are sourced from “hundreds of thousands” of real-world environments built largely from MiniMax’s own internal company tasks. Coding training in particular covers 10+ programming languages (Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart, Ruby) over 200,000+ real-world environments, with the data recipe explicitly covering the full development lifecycle (0→1 design, 1→10 development, 10→90 iteration, 90→100 review/testing). A trained behavior worth noting: M2.5 develops a Spec-writing tendency — before writing code it actively decomposes and plans features, structure, and UI from an “experienced software architect” perspective.

Search/tool-calling training is evaluated on RISE (Realistic Interactive Search Evaluation), MiniMax’s internal benchmark built around real expert questions that require multi-step browsing across information-dense webpages, layered on a Playwright tool suite over the WebExplorer agent framework. Office training partners with senior finance / law / social-science professionals who design requirements, define standards, and contribute to data construction; the eval framework (GDPval-MM, built on the open GDPval test set) uses LLM-as-judge to pairwise-compare full trajectories on deliverable quality, agent professionalism, and token cost.

The MiniMax Agent product is presented as both the deployment surface and the data source: Office Skills (typed Word/PowerPoint/Excel handlers) auto-load by file type in MAX mode, and users compose Skills with domain SOPs into reusable Experts — yielding a closed loop where production agent usage continuously generates verifiable training trajectories.

Results

Coding (the headline axis):

SWE-Bench Verified 80.2% (Claude Code scaffolding, default-prompt overridden, averaged over 4 runs); also reported on Droid scaffold 79.7 (vs Opus 4.6’s 78.9) and OpenCode scaffold 76.1 (vs 75.9).
Multi-SWE-Bench 51.3% — explicitly reported as surpassing Claude Opus 4.6.
VIBE-Pro (MiniMax’s internal extended VIBE benchmark): “on par with Opus 4.5”.
Note: the HF model card’s evaluation results widget lists SWE-Bench Verified resolved at 75.8 (marked with an asterisk), lower than the headline 80.2 — likely a different scaffold/run setting.

Search and tool-calling:

BrowseComp 76.3% (with context management; uses WebExplorer agent framework, discards all history when token usage exceeds 30% of max context).
BFCL multi-turn 76.8% — reported as “industry-leading”, an enormous +39.4 jump from M2.1’s reported number, and (per the Labonne writeup) +13 points over Opus 4.6 on the multi-turn split specifically.
~20% fewer rounds than M2.1 across BrowseComp / Wide Search / RISE.

Efficiency:

100 tok/s native serving for M2.5-Lightning; SWE-Bench Verified end-to-end runtime 22.8 min/task (37% faster than M2.1’s 31.3 min); ~3.52M tokens per task on SWE-Bench Verified.
$1/hour at 100 tok/s,$ 0.30/hour at 50 tok/s; output-price ratio 1/10–1/20 of Opus / Gemini 3 Pro / GPT-5.

Office work:

GDPval-MM internal pairwise eval: 59.0% average win rate vs mainstream models.

Reasoning / knowledge benchmarks (appendix, less foregrounded):

AIME25 86.3 (vs Opus 4.6 95.6, Gemini 3 Pro 96.0, GPT-5.2 98.0) — clearly behind.
GPQA-D 85.2 (vs Opus 4.5 87.0, Opus 4.6 90.0) — close.
HLE-without-tools 19.4 (vs Opus 4.6 30.7, GPT-5.2 31.4) — well behind.
SciCode 44.4 (vs Opus 4.6 52.0) — behind.
IFBench 70.0, AA-LCR 69.5 — competitive.

Independent reception (Labonne / OpenHands Index, surfaced via web search):

OpenHands placed M2.5 4th overall, behind only Claude Opus 4.6, Claude Opus 4.5, and GPT-5.2 Codex — i.e. M2.5 sits at the open-weights frontier of the coding-agent Pareto.
Early independent reports flag occasional sloppiness (wrong-branch pushes, missed formatting instructions) — “strong but occasionally sloppy”.

Why it’s interesting

For Luma this is the canonical “agentic frontier model” datapoint of Q1 2026: same architecture family as M2/M2.1 already on Agentic Software Engineering and Tool-Use Agents, but the headline gain (37% faster, Opus-comparable speed at 1/10 the cost) is attributed cleanly to RL scaling, not architecture. The Forge framework, CISPO + process-reward + completion-time-shaped reward, and the company-internal data flywheel (30% of internal tasks auto-completed; 80% of new code model-generated) sketch a recipe that competes with closed-source frontier models on agentic axes specifically — and arrives ahead of MiniMax’s promised separate technical blog on RL scaling, which when it lands will likely be the primary technical artifact to file.

The cost framing is striking and specific: “four M2.5 instances running continuously for an entire year for $10,000” is the kind of envelope that makes long-running agent deployments (continuous monitoring, asynchronous research agents, multi-instance creative iteration loops) economically tractable. The BFCL multi-turn 76.8% — +39.4 over M2.1 and +13 over Opus 4.6 — is large enough to deserve independent reproduction; if it holds it’s the strongest open-weights tool-calling result on record.

The interesting open question is what’s actually scaling: the 230B/10B-active architecture is fixed across M2 → M2.1 → M2.5, so the recipe attributes nearly all of the three-and-a-half-month improvement curve to (a) more / better training environments, (b) Forge throughput, and (c) reward shaping. If true, this is a clean datapoint for the “Reasoning RL at frontier scale” thesis without architectural confounds.