MiniMax has launched M3.1 Flash, an agentic model priced at $0.15 per million input tokens. The company confirmed that Space Bunny, which processed over 200 trillion tokens previously, was an earlier build of this same model. Developers now have access to near-frontier agentic scores at a flash-tier cost, with model weights scheduled for release in late October 2026.
Specs and availability
The system features 428 billion total parameters, with approximately 23 billion active in its mixture-of-experts configuration. It supports a context window of one million tokens. The model is currently available via the MiniMax API, specifically on the M Plan and MiniMax Code services. There is no information on the hardware required for self-hosting, though weights are due on October 23.
MiniMax reports performance as near-frontier for agentic office tasks, mid-pack for difficult coding, and top-tier on visual reasoning. It currently leads the Automation-Bench v1.0.6 chart with a score of 62.2. Conversely, its performance on Terminal Bench 4.0 is 38.1, which trails Opus 5 at 51.8.
Technical details
M3.1 Flash is a multimodal mixture-of-experts model designed for coding, agentic office work, and visual reasoning. It accepts text, image, and video input, returning text. The architecture includes sparse attention and a native visual encoder.
Reasoning is always active and cannot be disabled. Users must select from five effort levels: Low, Medium, High, Xhigh, and Max. MiniMax recommends Max as the default setting. Requests attempting to disable thinking via `thinking.type=disabled` or `effort=none` will return an error.
Integrations that previously parsed inline `<think>` tags from M3 must be updated. Reasoning output now appears in a separate `reasoning_content` field, which may be absent depending on the configuration.
Benchmark results
All scores below are provided by MiniMax. No independent replication exists yet.
On the CoWork set, M3.1 Flash leads Automation-Bench v1.0.6 at 62.2. Opus 5 scores 50.3, while GPT-5.6 Sol scores 45.8. It ties Opus 5 on BankerToolBench at 76.2 and tops LOCA-Bench 256K at 60.3. Agents’ Last Exam scores 29.6.
Coding results are mixed. DeepSWE v1.1 scores 73.7, just above GPT-5.6 Sol at 73.0. DeepSeek-V4.1-Flash-1017 scores higher at 77.0. Terminal Bench 4.0 is the weakest result at 38.1, while SWE Refactor Bench sits at 23.0 compared to 31.0 for Opus 5.
On vision tasks, BabyVision with tools reaches 91.0 and CharXiv RQ reaches 82.5. HLE without tools scores 43.4.
Spot checks hold up. The Kimi K3 model card reports DeepSWE 67.5 and BabyVision 85.7, matching MiniMax’s table. One inconsistency remains: MiniMax’s cost chart plots Automation-Bench max effort at 62.8, while its table says 62.2.
Cost per task
MiniMax frames value as cost per task rather than per token. On its Automation-Bench chart, all five effort levels sit below $0.05 per task. Scores range from 52.8 at low effort to 62.8 at max. A full 152-task Agents’ Last Exam run cost about $111 at 29.6. On DeepSWE, DeepSeek V4.1 Flash-1017 plots both higher and cheaper at max effort.
Deployment and pricing
The model is live on platform.minimax.io as `MiniMax-M3.1-Flash`. MiniMax supports Anthropic-compatible and OpenAI-compatible access, including Chat Completions and Responses. It is also available through the M Plan and the MiniMax Code desktop app.
Pricing is a flat $0.15 input and $0.60 output per million tokens. Cache reads cost $0.015 per million. There is no long-context surcharge. MiniMax M3 costs $0.30/$1.20 up to 512K input and doubles above that. Per token, Flash is half of M3’s standard rate. MiniMax claims one-third the cost of M3 but has not detailed the calculation.
On speed, the product page lists about 150 tokens per second. The announcement claims three times the speed of M3. Weights arrive October 23; the license is not yet disclosed. For scale, M3’s MXFP8 build is about 440 GB.
Comparison with competitors
This table compares M3.1 Flash with DeepSeek V4.1 Flash, Kimi K3, and GLM-5.3.
- Parameters: M3.1 Flash is 428B total / ~23B active. DeepSeek V4.1 Flash is 552B total / 8B prefill, 16B decode. Kimi K3 is 2.8T total / 104B active. GLM-5.3 is ~744B total / ~40B active.
- Context: All four models support one million tokens.
- Inputs: M3.1 Flash accepts text, image, and video. DeepSeek and Kimi accept text and image. GLM accepts text only.
- Weights: M3.1 Flash weights are due October 23 with no disclosed license. DeepSeek and GLM offer open weights. Kimi K3 uses the Kimi K3 License.
- Pricing (per 1M tokens): M3.1 Flash is $0.15 input / $0.60 output. DeepSeek is $0.30 / $1.20 peak (half off-peak). Kimi is $3.00 / $15.00. GLM is $1.40 / $4.40.
- Automation-Bench v1.0.6: M3.1 Flash (62.2), DeepSeek (54.8), Kimi (46.7), GLM (48.8).
- DeepSWE v1.1: M3.1 Flash (73.7), DeepSeek (74.2), Kimi (67.5), GLM (66.9).
- Terminal Bench 4.0: M3.1 Flash (38.1), DeepSeek (31.2), Kimi (12.6), GLM (37.9).
- BabyVision (w/ tools): M3.1 Flash (91.0), DeepSeek (89.6), Kimi (85.7), GLM (not reported).
Limitations
Every benchmark is vendor-reported, and harness choice shifts results. Automation-Bench numbers do not transfer across sources. Kimi K3’s card reports 30.8 on a 600-task public subset, while MiniMax lists 46.7.
Opus 5 leads by 13.7 points on Terminal Bench 4.0 and 14.1 on NL2Repo-Bench. Always-on reasoning also sets a token and latency floor for simple calls. Self-hosting details remain unknown until October 23.
