ZICQ
中 Log in / Sign up
Newsroom Open Source AI #MiniMax M2 #Open-Source Model #AI Agents #Tool Use #Inference Cost

MiniMax Open-Sources M2 and Agent for Low-Cost Tool Use

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:MiniMax has launched and open-sourced M2, a model designed around agent workloads, while making a free API and an Agent product available. According to the company, the model delivers leading-tier coding, tool use, and deep-search performance at 8% of Claude Sonnet's price and at twice the speed. If the pricing units are directly comparable, that headline rate implies roughly a 12.5-fold reduction in unit token cost. For agent engineering, however, nominal token price is only one component of c


Core Release

MiniMax has launched and open-sourced M2, alongside an Agent product and a free API. The company positions M2 as an agent-first model focused on coding, tool use, and deep search, claiming performance at 8% of Claude Sonnet's price and twice its speed.

If input, output, and cached-token pricing are directly comparable, an 8% rate implies roughly one-twelfth the unit token cost, or a theoretical 12.5-fold price advantage. That matters more for agent systems, which repeatedly invoke a model, execute tools, and ingest intermediate results, than a conventional single-turn benchmark would suggest.

What Agent-First Actually Means

Agent-first describes the intended interaction between a model and an execution system; it does not, by itself, prove the use of a new training paradigm. A typical agent loop requires the model to emit structured tool calls, the runtime to validate and execute them, the resulting observations to be appended to the context, and the model to decide whether to continue, repair an invalid call, or return a final answer.

Four engineering factors will determine M2's real-world performance:

  1. Tool-call reliability: Schema adherence, invalid-argument rates, parallel-call capability, and recovery after tool failures.
  2. Long-horizon state management: Maintaining goals, constraints, and evidence across multi-step search, code modification, and task handoffs.
  3. Context economics: Tool results expand the prompt used in every subsequent turn. For a conventional Transformer backend, per-sequence KV-cache memory scales approximately as 2 × layers × sequence length × KV heads × head_dim × bytes × batch. GQA, MLA, MQA, paging, and prefix reuse alter the real footprint. Since the announcement does not disclose M2's architecture, this formula cannot be used to estimate its exact hardware requirements.
  4. Deep-search quality: Retrieving relevant material is only the first step; production systems also need source freshness, cross-verification, traceable citations, and resistance to prompt injection in retrieved pages.

Cost and Speed Trade-offs

A model's advertised token price is not the same as the cost of an agent task. A more useful calculation is:

Total cost per task ≈ Σ(input-token cost + output-token cost) + tool and retrieval fees + retry cost + orchestration cost

If M2 needs more reasoning turns, produces longer plans, or issues more invalid tool calls to complete the same objective, part of the 12.5-fold unit-price advantage will be consumed by additional calls. Prefix caching helps only when the system prompt, tool definitions, and historical context remain stable. Dynamic web content, search results, and continuously expanding traces can reduce the hit rate.

The claimed 2x speed also needs to be decomposed. Time to first token, decoding throughput, concurrent serving throughput, and end-to-end task latency are different measurements. Search, browsers, code sandboxes, and external APIs may contribute more wall-clock time than model inference. Production evaluation should therefore measure latency per successfully completed task rather than generation speed alone.

Open weights provide self-hosting, data control, auditability, and opportunities for adaptation. They also transfer GPU procurement, inference-stack compatibility, quantization error, KV-cache management, and model upgrade work to the operator. Whether self-hosting is economical depends on utilization and peak concurrency, not token price alone.

Performance Claims and Evaluation Gaps

The available material does not provide absolute scores, parameter counts, active parameters, context length, license terms, weight locations, hardware configurations, sampling parameters, or a complete evaluation table. As a result, the claims of top-tier performance, 8% pricing, and 2x speed cannot yet be independently verified.

Coding should be tested on real repositories, with patch acceptance rate, test pass rate, changed-line scope, and regression rate as primary metrics rather than relying only on HumanEval-style function completion. Tool-use evaluation should record tool-selection accuracy, argument validity, retries, and unrecoverable failures. Deep search should examine citation accuracy, source coverage, answer falsifiability, and resistance to malicious instructions embedded in retrieved content.

Any comparison with Claude Sonnet must pin the exact model version and use identical system prompts, tool sets, token budgets, reasoning settings, caching policies, and retry limits. If one model wins only with a larger reasoning budget or more external tools, the difference should not be attributed entirely to the base model.

Deployment Recommendations

  1. Run a same-stack A/B test: Keep the orchestration framework fixed and replace only the model interface when comparing Claude Sonnet with M2.
  2. Price successful tasks: Track input, output, tool, retry, search, and sandbox costs per completed task, alongside p50 and p95 end-to-end latency.
  3. Test context depth in layers: Evaluate short tasks, trajectories accumulating hundreds of thousands of tokens, and large code repositories separately to distinguish model limits, KV-cache pressure, and truncation policies.
  4. Measure quantization loss: If using FP8 or INT4, verify structured outputs, code patches, and tool arguments; otherwise quality degradation may be misattributed to the model.
  5. Apply least-privilege controls: Put code execution, file writes, web access, and credential use behind sandboxing, scoped permissions, and human approval. Treat all retrieved and tool-generated content as untrusted input.
  6. Verify the openness boundary: Confirm whether weights, training code, license terms, and commercial-use rights are available together. A weights-only release with limited reproducibility offers less ecosystem value than a complete open-source project.

The main appeal of MiniMax M2 is the combination of an open model, a free API, and an agent product centered on low-cost execution. If the price and speed advantages survive testing under identical conditions, M2 could lower the barrier to building coding and deep-research agents. Until the architecture and benchmark details are disclosed, it is better treated as a high-value candidate for hands-on evaluation than a verified performance leader.


Source: MiniMax Blog (2025-10-27)

— END —

Tags: #MiniMax M2 #Open-Source Model #AI Agents #Tool Use #Inference Cost

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…