MiniMax M3 and Qwen3.7 Plus on Critique: Coding Benchmarks and a Two-Week M3 Welcome Price

June 2026 is another “two launches, one economics story” week. MiniMax shipped M3 as an open-weight multimodal foundation model with a million-token context window and sparse attention (MSA). Alibaba positioned Qwen3.7 Plus as a multimodal agent model — not merely a vision upgrade, but a single loop that can move between terminal output, UI screenshots, and tool calls. Critique’s job is not to repeat vendor keynote charts. It is to wire the models on the surfaces that spend review credits — lead synthesis, specialist passes, Remedy execution — and price them so the median PR can afford frontier-class coding without defaulting to Opus on every file. Repo chat stays on its own free roster.

PART ONE — MINIMAX M3: WHAT CHANGED VS M2.7

MiniMax M2.7 was the self-evolving agent lane Critique teams already knew: strong SWE-Bench Verified numbers on the Hugging Face card (80.2% in our catalog snapshot), 56.22% on SWE-Bench Pro in MiniMax’s M2.7 launch materials, and 57.0% on Terminal Bench 2 in the same generation. M3 is a architectural step, not a point release. MiniMax describes M3 as natively multimodal from step zero, trained with interleaved text-image (and video-capable) data, and served with up to 1M tokens via MSA — sparse attention that cuts per-token compute at long context (they quote ~1/20 the cost of the prior generation at 1M tokens, with large prefill/decode speedups in their launch post).

MiniMax’s own positioning is aggressive: on SWE-Bench Pro they report beating GPT-5.5 and Gemini 3.1 Pro and approaching Claude Opus 4.7; on BrowseComp they report 83.5 vs 79.3 for Opus 4.7. Anthropic has since shipped Opus 4.8, so the correct buyer question is not “does M3 beat last month’s Opus row?” but “does M3 clear my quality bar at 1/10th the credit burn?” For many repos, the answer is now plausibly yes on coding-and-tools workloads — with the usual caveat that long-horizon autonomy tests in vendor blogs (12-hour paper reproduction, 24-hour CUDA kernel search) are demonstrations, not guarantees on your monorepo.

PART TWO — QWEN3.7 PLUS: MULTIMODAL AGENT VS QWEN3.6 PLUS

Qwen3.7 Plus is already the Alibaba workhorse in Critique’s runtime catalog: specialist passes, cheap-volume stacks, and the default free Remedy chat model id (`qwen/qwen3.7-plus`). The June 2026 launch framing from Alibaba and press coverage is different from “slightly better Qwen3.6.” Qwen3.7 Plus is pitched as a multimodal agent foundation — text, image, and video inputs — with stronger computer-use and terminal scores than the prior Plus generation.

Practical guidance: keep Qwen3.7 Plus when the PR carries visual signal — Figma exports, failing UI screenshots, chart-heavy docs, short screen recordings. Escalate to Qwen3.7-Max (6 credits on Critique) when the job is terminal-only, maximum reasoning depth, and you can afford the step up. For the lowest-burn text lanes on review, MiMo v2.5 and Ling still anchor the sub-1cr chart from our May catalog essay.

PART THREE — CRITIQUE SURFACES AND PRICING

PART FOUR — M3 VS OPUS 4.8 AND COMPOSER 2.5 (BUYING FRAME)

Three questions land in Discord every launch week. The charts above hold the scores — here is the buying frame without mixing benchmarks.

PART FIVE — SOURCES AND REPRODUCIBILITY