MiniMax-M2.7 speed and noncommercial limits

Results and use limits

I ran MiniMax-M2.7 (229B MoE) on two RTX PRO 6000 Blackwell 96GB GPUs. Ten samples read from the video gave average TG of 71.9 t/s and a maximum of 105.9 t/s.

The modified-MIT terms at the time of writing require prior written permission from MiniMax for commercial use. I kept this as a noncommercial benchmark and deleted the model. This is an LLM evaluation and inference-infrastructure record, not a customer deployment test.

It remains a candidate for personal prompt and tool-call checks. DGX Spark and Mac Studio 128GB were not tested; the run used dual Blackwell 96GB.

Context

I read 10 slot print_timing samples from the 14-minute uncut YouTube video. Separate inference logs were not saved.

The task was to hand Zed Editor’s AI Agent a WordPress-like Django blog specification and see how far it could autonomously build it.

Video link: https://www.youtube.com/watch?v=4QIc4A5Yly4

Test Environment

Item	Specification
GPU	2x NVIDIA RTX PRO 6000 Blackwell 96GB
CPU	AMD EPYC 9175F
RAM	768GB
Engine	ik_llama.cpp (ubergarm fork)
Quantization	smol-IQ3_KS
Context	262,144 tokens
KV buffer	CUDA0 32,768 MiB + CUDA1 30,720 MiB
Prompt cache	8,192 MiB
Graph splits	3

Benchmark Results

Token Generation

Phase	KV cache	Speed
Initial	~2.7k tokens	105.9 t/s
Early average	under 25k tokens	77.5 t/s
Mid average	27k-47k tokens	64.7 t/s
Late average	54k-56k tokens	54.3 t/s

Mean: 71.9 t/s
Max: 105.9 t/s
Min: 53.9 t/s
Degradation: -49% from 2.7k to 56.4k context

Prompt Processing

Condition	Speed
Cold start (2,493 tokens)	4,067 t/s
Cache-hit average	911 t/s

10 Sample Points

KV cache	TG
2.7k	105.9 t/s
18.3k	79.2 t/s
19.1k	78.8 t/s
21.2k	77.1 t/s
23.7k	75.0 t/s
27.2k	71.4 t/s
38.4k	63.8 t/s
46.6k	58.9 t/s
53.9k	54.7 t/s
56.4k	53.9 t/s

Sampled speed declined as context grew. Generation continued at 56k. Sessions beyond 100k were not measured and would need task-splitting checks.

What Was Generated

Given a WordPress-like Django blog/CMS specification, Zed Agent produced the following in an uncut session of about 14 minutes:

Full Django config
8 apps: accounts, content, core, media, comments, navigation, seo, taxonomy
Model, admin, and app config for each app
Test suite
Gitea issue creation API integration
More than 50 files

The agent noticed and corrected an edit_file path prefix problem during execution. No independent QA result for the complete generated application is presented.

Session screenshots

MiniMax-M2.7 server initialization and model loading screen — Startup — model loading and initialization

Zed Agent generating Django configuration and structure — Early phase — generating Django configuration and structure

Mid-session screen with multiple Django apps scaffolded — Mid-session — multiple apps taking shape

Final state with 50 plus files and test suite generated — Late phase — 50+ files and a test suite in place

Differences from the M2.5 run

The MiniMax-2.5 run maintained 34–37 t/s on one 96GB GPU with Expert Offload, including relatively stable TG at longer context.

M2.5 used one GPU with Expert Offload; M2.7 used dual Blackwell with smol-IQ3_KS. These are different conditions. Subjectively, M2.7 appeared to have more headroom in tool use and output.

Token usage was about 66k here. I remember about 122k for M2.5, so the roughly half-sized comparison retains that uncertainty.

Licensing determined the decision against routine use. I followed community discussion but limited the work to this benchmark.

Use conditions for M2.5 and M2.7

M2.7’s modified-MIT does not freely allow commercial use.

When I read M2.5’s terms in March 2026, I did not find the same restriction. Fine-tuning M2.5 is another possible route, with the applicable terms still needing review at use time.

Candidate uses are benchmarking, study and strictly personal local checks.
I have not introduced it into business, customer or internal commercial workflows.
It was not adopted as a default local model.

I could not determine how far the restriction extends to outputs or secondary artifacts. That uncertainty also ruled out routine use.

Measurement Method

The source data is the video because separate inference logs were not saved.

Claude extracted PP/TG from selected frames, so reading errors are possible. The original displayed values are in the video.