MiMo-V2.6-Flash Official X Release Thread: Flash's Benchmark Positioning and Dual-Model Strategy

One-sentence takeaway

The official post positions MiMo-V2.6-Flash as an omnimodal model alongside Pro, emphasizing scaled reinforcement learning, coding/general Agent/cybersecurity/visual Agent capabilities, and open reproduction; the attached image gives Flash's individual benchmark scores but does not disclose the sample size, complete harness, hardware, prompts, or decoding parameters, so these figures can only be treated as a vendor baseline.

Use cases

• Tasks suitable for assessment: Official capability positioning for coding agents, general agents, cybersecurity agents, and visual agents; assessing the product division between Flash and Pro and their open-source direction.

• Tasks unsuitable for extrapolation: Real-world business success rates, latency, throughput, cost, stability, or performance under different toolchains or inference parameters; Pro-exclusive claims must not be attributed to Flash.

• Applicable model version: MiMo-V2.6-Flash. This post does not use the checkpoint name MiMo-V2.6-Flash-RL.

• Test environment or client: Xiaomi MiMo's official X post and its attached image; the evaluation hardware, client, API endpoint, and harness are not specified.

• Inference tier and parameters: Not specified.

Evaluation method

The only visible evaluation material is a table in the image attached to the original post (captioned “Table 3 Comparison of MiMo-V2.6 with previous-generation and frontier models on agentic benchmarks”). The table groups benchmarks into Code Agent, General Agent, Cybersecurity, and Visual Agent, and lists MiMo-V2.6 Pro, MiMo-V2.6 Flash, MiMo-V2.5 Pro, Claude Opus 5, GPT-5.6 Sol, and Fable 5 across the columns.

The original post says that the two models were advanced through scaled reinforcement learning and claims stronger coding, computer use, 3D reasoning, and creative capabilities; the visible thread also says that API pricing for both Pro and Flash remains unchanged from V2.5, and that Pro, Flash, the technical report, 7K+ RL task environments, an end-to-end RL framework, and composable mini-harnesses are being open-sourced. These are vendor release statements, not a complete disclosure of the evaluation method.

The source does not specify the sample size for each benchmark, task-sampling rules, evaluation date, complete prompts, few-shot setup, tool permissions, decoding parameters, hardware, number of repetitions, scoring scripts, confidence intervals, or failed samples.

Key results

The table below transcribes only the MiMo-V2.6 Flash column from the attached image; — means that no value was provided in the image.

Raw data

• Original post positioning: Introducing Xiaomi MiMo-V2.6 — Pro & Flash.; Two omnimodal models, advancing through scaled reinforcement learning.

• Original post capability statement: Stronger coding, computer use, 3D reasoning and creative capabilities.

• Original post open-source claim: Open model weights, technical report, RL environments and training code.

• Flash-related statements in the visible thread: API pricing unchanged from V2.5, for both Pro and Flash; We’re open-sourcing Pro and Flash ... 7K+ RL task environments ... composable mini-harnesses.

• Image attached to the original post: https://pbs.twimg.com/media/HSxK_3caUAAsfBl?format=jpg&name=medium. In the image's Flash column, CyberGym is 95.1, Terminal Bench 2.1 is 87.6, and MiMo Visual Coding is 71.5; the remaining items in the table above are also listed in full.

• Explicit version boundary: In the original post, Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks and Pro scores 46 ... both clearly refer to Pro, not Flash results.

Conclusions and limitations

• Vendor claim: Flash was released as an omnimodal model alongside Pro, with a single Agent benchmark table showing its coding, general Agent, cybersecurity, and visual Agent scores; the official thread also emphasizes unchanged API pricing and open RL resources.

• Independent measurement: None provided. This page contains no independent rerun, third-party harness, or personal experience data.

• Data interpretation limits: The table is a snapshot from one official release and is insufficient to infer overall rankings or average cross-task capability. The score definitions for different benchmarks are also not uniformly explained.

• Comparison limits: The comparison models and Flash in the attached image may not have used the same time period, hardware, tools, prompts, or harness; differences in the table cannot be interpreted as strict causal advantages.

• Model boundary: Do not rewrite data from MiMo-V2-Flash, MiMo-V2.5-Pro, MiMo-V2.6-Pro, or MiMo-V2.6-Pro-UltraSpeed as Flash data; this page records only the MiMo-V2.6 Flash column shown in the image.

Reproduction notes

1. Open the original X post, confirm that the author is @XiaomiMiMo, and inspect the MiMo-V2.6 Flash column in the image attached to the original post.

2. Transcribe the scores item by item by category and benchmark name; keep the image's — as “not provided” and do not rewrite it as 0.

3. To conduct an independent rerun, obtain the task-set version, complete harness, prompts, tool configuration, decoding parameters, hardware, sample size, and scoring scripts separately; the original post itself is insufficient to reproduce the experiment.

4. Store independent results in a separate column from the vendor values in this article, and record the actual model identifier, client, run time, and failed samples; do not use independent results to replace the official table.