Breaking

MiniMax M3 Added to Agent Arena: An Open-Weight Model Supporting 1M Context

June 12, 2026 at 18:41 EDT

Foundation Models
AI Agents
Open Source

MiniMax M3, an open-weight foundation model from China's MiniMax, has been added to Agent Arena, a real-world agent evaluation platform. On the platform, the model is given web search, filesystem and terminal tools and is evaluated by completing complex workflows such as writing code, creating slide decks, researching the web, building apps and analyzing documents. A leaderboard score for M3 is said to be coming soon.

June 1, 2026 · MiniMax

MiniMax M3: An Open-Weight Bid for the Agentic Frontier

A 428B-parameter MoE model claiming to be the first open-weight release to unite frontier coding, a 1M-token context, and native multimodality — now entered into real-world agent evaluation.

428B

Total params (~23B active, MoE)

Token context (512K guaranteed)

59%

SWE-Bench Pro (vendor claim)

~2,000

Tool calls in 24h runs (reported)

MiniMax Sparse Attention (MSA) · Speed vs. prior generation

KV-block selection replaces full attention to slash long-context compute. Columns scale to each speedup multiplier.

1×

Baseline
(prior gen)

9.7×

Faster
prefill

15.6×

Faster
decode

Net effect: roughly 1/20 the cost at 1M tokens — making long-horizon agents with many tool calls economically feasible.

Token pricing · per 1M tokens

Standard

$0.60 in / $2.40 out

Promo

$0.30 in / $1.20 out

Positive reports

Strong on coding & agent tasks (e-commerce builds, complex scenario tests)
1M context useful for long-running execution
24-hour continuous runs, ~2,000 tool calls

Cautions

"Unpredictable," hallucination-prone, verbose, stops mid-task
Weaker at abstract reasoning
GGUF quantization loses MSA benefit & bloats KV cache

The open question

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.