---
type: "post"
id: "mimo-v2-6-pro-training-notes-analysis-7aef31"
slug: "mimo-v2-6-pro-training-notes-analysis-7aef31"
title: "MiMo-V2.6 Pro Training Notes"
description: "MiMo-V2.6 Pro is a useful signal that frontier open-weight gains may increasingly come from agent-oriented post-training systems rather than from conspicuously novel attention architectures."
retrieval_nugget: "MiMo-V2.6 Pro is a useful signal that frontier open-weight gains may increasingly come from agent-oriented post-training systems rather than from conspicuously novel attention architectures."
published_at: "2026-09-22"
updated_at: "2026-09-22"
record_date: "2026-09-22"
date_kind: "published_at"
topics: []
source_urls: ["https://sebastianraschka.com/blog/2026/mimo-v2-6-pro-architecture-training-notes.html"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:mimo-v2-6-pro-training-notes-analysis-7aef31"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:d762cb37d60e2e177e96bbd70379355969f0d2780f53a6fbed719f991c4fe93e","reviewed_at":"2026-09-22T18:00:52.204Z","source_evidence_count":1,"verified_claim_count":4,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":4,"implication_count":2,"watch_condition_count":2,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"MiMo-V2.6 Pro is a useful signal that frontier open-weight gains may increasingly come from agent-oriented post-training systems rather than from conspicuously novel attention architectures.","observed_facts":[{"text":"The supplied analysis describes GQA with sliding-window attention using a 128-token window.","source_urls":["https://sebastianraschka.com/blog/2026/mimo-v2-6-pro-architecture-training-notes.html"]},{"text":"It reports held-out-harness DeepSWE pass@1 improving from approximately 50% to 66%.","source_urls":["https://sebastianraschka.com/blog/2026/mimo-v2-6-pro-architecture-training-notes.html"]},{"text":"It cites RL batches of 1,568 prompts times 16 rollouts, totaling 25,088 trajectories.","source_urls":["https://sebastianraschka.com/blog/2026/mimo-v2-6-pro-architecture-training-notes.html"]},{"text":"It says the grader evaluates execution traces in addition to correctness.","source_urls":["https://sebastianraschka.com/blog/2026/mimo-v2-6-pro-architecture-training-notes.html"]}],"mechanism":"The mechanism is a tighter learning loop: diverse execution environments expose policy brittleness, trace-aware grading supplies denser information about how an agent reached an answer, and large rollout batches create more training signal per update.","why_now":"This matters now because the report couples those operational changes with a leading open-weight result while describing an otherwise familiar attention design.","implications":["The implication is that teams building coding or tool-using agents should treat harness diversity and trajectory evaluation as model-development assets, not merely benchmark infrastructure.","Architecture comparisons should not be used as proxies for the quality of a model’s data, grading, and post-training system."],"evidence_boundary":"The evidence is a secondary technical analysis summarizing a MiMo technical report, not an independently reproduced ablation study; it does not establish the causal contribution of each intervention.","watch_conditions":["Watch whether MiMo or independent evaluators publish controlled ablations separating agent-task mix, cross-harness exposure, trace-aware grading, and RL-scale effects.","Watch whether cross-harness gains persist on unrelated agent benchmarks and real tool-use workloads."],"related_records":[],"new_branch_reason":"No prior New Runtime record was supplied; this begins a distinct record on agentic post-training infrastructure and evaluation design."}
routes: {"html":"https://newruntime.com/posts/mimo-v2-6-pro-training-notes-analysis-7aef31/","markdown":"https://newruntime.com/posts/mimo-v2-6-pro-training-notes-analysis-7aef31.md","json":"https://newruntime.com/posts/mimo-v2-6-pro-training-notes-analysis-7aef31.json"}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
---

# MiMo-V2.6 Pro Training Notes

## Retrieval answer

MiMo-V2.6 Pro is a useful signal that frontier open-weight gains may increasingly come from agent-oriented post-training systems rather than from conspicuously novel attention architectures.

MiMo-V2.6 Pro is a useful signal that frontier open-weight gains may increasingly come from agent-oriented post-training systems rather than from conspicuously novel attention architectures.

The supplied analysis describes a conventional GQA architecture with sliding-window attention at a 128-token window, alongside a recipe that increases agent tasks and trains across multiple harnesses. It reports that DeepSWE pass@1 on held-out harnesses improved from approximately 50% to 66%. The author also describes replacing a simple correctness verifier with an agentic grader that examines execution traces, and cites RL batches of 1,568 prompts times 16 rollouts, or 25,088 trajectories.

The mechanism is a tighter learning loop: diverse execution environments expose policy brittleness, trace-aware grading supplies denser information about how an agent reached an answer, and large rollout batches create more training signal per update. The implication is that teams building coding or tool-using agents should treat harness diversity and trajectory evaluation as model-development assets, not merely benchmark infrastructure. This matters now because the report couples those operational changes with a leading open-weight result while describing an otherwise familiar attention design.

The evidence is a secondary technical analysis summarizing a MiMo technical report, not an independently reproduced ablation study; it does not establish the causal contribution of each intervention. Watch whether MiMo or independent evaluators publish controlled ablations separating agent-task mix, cross-harness exposure, trace-aware grading, and RL-scale effects.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)