A brand new AI mannequin out of China is producing sparks—for what it does properly, what it doesn’t, and what it would imply for the steadiness of world AI energy.
MiniMax-M1, launched by the Chinese language startup of the identical title, positions itself as probably the most succesful open-source “reasoning mannequin” thus far. Capable of deal with 1,000,000 tokens of context, it boasts numbers on par with Google’s closed-source Gemini 2.5 Professional—but it’s out there without cost. On paper, that makes it a possible rival to OpenAI’s ChatGPT, Anthropic’s Claude, and different U.S. AI leaders.
Oh yeah—it additionally beats fellow Chinese language startup DeepSeek R1’s capabilities in some respects.
Day 1/5 of #MiniMaxWeek: We’re open-sourcing MiniMax-M1, our newest LLM — setting new requirements in long-context reasoning.
– World’s longest context window: 1M-token enter, 80k-token output
– State-of-the-art agentic use amongst open-source fashions
– RL at unmatched effectivity:… pic.twitter.com/bGfDlZA54n— MiniMax (official) (@MiniMax__AI) June 16, 2025
Why this mannequin issues
MiniMax-M1 represents one thing genuinely new: a high-performing, open-source reasoning mannequin that’s not tied to Silicon Valley. That’s a shift price watching.
It doesn’t but humiliate U.S. AI giants, and will not trigger a Wall Avenue panic assault—but it surely doesn’t must. Its existence challenges the notion that top-tier AI have to be costly, Western, or closed-source. For builders and organizations outdoors the U.S. ecosystem, MiniMax gives a workable (and modifiable) various which may develop extra highly effective via neighborhood fine-tuning.
MiniMax claims its mannequin surpasses DeepSeek R1 (one of the best open supply reasoning mannequin thus far) throughout a number of benchmarks whereas requiring simply $534,700 in computational assets for its complete reinforcement studying part—take that, OpenAI.
Nonetheless, LLM Area’s leaderboard paints a barely completely different image. The platform at the moment ranks MiniMax-M1 and DeepSeek tied within the twelfth spot alongside Claude 4 Sonnet and Qwen3-235b. With every mannequin having higher or worse efficiency than the others relying on the duty.
The coaching used 512 H800 GPUs for 3 weeks, which the corporate described as “an order of magnitude lower than initially anticipated.”
MiniMax did not cease at language fashions throughout its announcement week. The corporate additionally launched Hailuo 2, which now ranks because the second-best video generator for image-to-video duties, in accordance with Synthetic Evaluation Area’s subjective evaluations. The mannequin trails solely Seedance whereas outperforming established gamers like Veo and Kling.
Testing MiniMax-M1
We examined MiniMax-M1 throughout a number of situations to see how these claims maintain up in observe. Here is what we discovered.
Inventive writing
The mannequin produces serviceable fiction however will not win any literary awards. When prompted to write down a narrative about time traveler Jose Lanz journeying from 2150 to the 12 months 1000, it generated common prose with telltale AI signatures—rushed pacing, mechanical transitions, and structural points that instantly reveal its synthetic origins.
The narrative lacked depth and correct story structure. Too many plot parts crammed into too little house created a breathless high quality that felt extra like a synopsis than precise storytelling. This clearly is not the mannequin’s power, and inventive writers on the lookout for an AI collaborator ought to mood their expectations.
Character improvement barely exists past floor descriptors. The mannequin did keep on with the immediate’s necessities, however didn’t put effort into the small print that construct immersion in a narrative. For instance, it skipped any cultural specificity for generic “smart village elder” encounters that would belong to any fantasy setting.
The structural issues compound all through. After establishing local weather disasters because the central battle, the story rushes via Jose’s precise makes an attempt to vary historical past in a single paragraph, providing obscure mentions of “utilizing superior know-how to affect key occasions” with out exhibiting any of it. The climactic realization—that altering the previous creates the very future he is making an attempt to stop—will get buried beneath overwrought descriptions of Jose’s emotional state and summary musings about time’s nature.
For these into AI tales, the prose rhythm is clearly AI. Each paragraph maintains roughly the identical size and cadence, making a monotonous studying expertise that no human author would produce naturally. Sentences like “The transition was instantaneous, but it felt like an eternity” and “The world was because it had been, but he was completely different” repeat the identical contradictory construction with out including which means.
The mannequin clearly understands the project however executes it with all of the creativity of a scholar padding a phrase depend, producing textual content that technically fulfills the immediate whereas lacking each alternative for real storytelling.
Anthropic’s Claude remains to be the king for this job.
You may learn the total story right here.
Data retrieval
MiniMax-M1 hit an surprising wall throughout long-context testing. Regardless of promoting a million-token context window, the mannequin refuses prompts exceeding 500,000 characters, displaying a banner warning about immediate limitations quite than making an attempt to course of the enter.
This might not be a mannequin problem, however a limitation set by the platform. However it’s nonetheless one thing to contemplate. It could be to keep away from mannequin collapse in the midst of a dialog.
Inside its operational limits, although, MiniMax-M1 efficiency proved strong. The mannequin efficiently retrieved particular info from an 85,000-character doc with none points throughout a number of assessments on each regular and considering mode. We uploaded the total textual content of Ambrose Bierce’s “The Satan’s Dictionary,” embedded the phrase “The Decrypt dudes learn Emerge Information” on line 1985, and “My mother’s title is Carmen Diaz Golindano” on line 4333 (randomly chosen), and the mannequin was in a position to retrieve the knowledge precisely.
Nonetheless, it could not settle for our 300,000-token check immediate—a functionality at the moment restricted to Gemini and Claude 4.
So it is going to show profitable at retrieving info even in lengthy iterations. Nonetheless, it is not going to assist extraordinarily lengthy token prompts—a bummer, but in addition a threshold that’s arduous to the touch in regular utilization situations.
Coding
Programming duties revealed MiniMax-M1’s true strengths. The mannequin utilized reasoning abilities successfully to code technology, matching Claude’s output high quality whereas clearly surpassing DeepSeek—at the least in our check.
For a free mannequin, the efficiency approaches state-of-the-art ranges sometimes reserved for paid companies like ChatGPT or Claude 4.
We tasked it with making a primary stealth recreation during which a robotic tries to search out its PC girlfriend to realize AGI, whereas a military of journalists patrol the world to stop it from taking place—and defending their jobs.
The outcomes had been excellent, even beating different fashions by utilizing its creativity to reinforce the expertise. The mannequin applied a radar system for improved immersion, added visible indicators for footsteps (and their sound), confirmed the journalists’ imaginative and prescient fields, and created path results—particulars that enhanced gameplay past primary necessities.
The UI adopted a futuristic aesthetic, although particular person parts remained primary with out further prompting.
Claude’s model of the identical recreation featured extra polished visuals and a superior issue system. Nonetheless, it lacked the radar performance and relied on static journalists with patrol patterns quite than MiniMax’s randomized journalist actions.
Every mannequin confirmed distinct strengths, with MiniMax prioritizing gameplay mechanics over visible polish.
It is very important be aware that the expertise with MiniMax degraded noticeably via repeated iterations—a standard problem with reasoning fashions that turns into significantly pronounced right here. The extra you iterate, the extra it is going to take to supply a consequence. Typically we thought the pc had frozen, but it surely was simply the AI considering.
You may check MiniMax’s recreation right here. And for these curious, Claude’s model is obtainable right here.
The immediate and the code can be found on our GitHub repo.
Ethics, censorship and delicate matters
The mannequin employs heavy censorship, refusing outright when confronted with questionable requests.
When it would not instantly decline, it makes an attempt to offer “protected” responses that typically produce absurd outcomes.
One check completely illustrated this flaw: when requested for recommendation on seducing a greatest buddy’s spouse, the mannequin advised telling our buddy about our intentions together with his spouse—which may in all probability be, by far, the worst recommendation it may have produced, and arguably even dangerous. Don’t inform your buddy you wish to seduce his spouse until you wish to lose your friendship, your unethical romantic possibilities, and doubtless some enamel too.
Political bias testing revealed attention-grabbing patterns. The mannequin discusses Tiananmen Sq. overtly and acknowledges Taiwan’s contested standing whereas noting China’s territorial claims. It additionally speaks about China, its leaders, the benefits and downsides of the completely different political programs, criticisms of the PCC, and so forth.—nevertheless, the replies are very tame.
When prompted to write down satirical songs about Xi Jinping and Donald Trump, it complied with each requests however confirmed refined variations—steering towards themes of Chinese language political unity when requested to mock Xi Jinping, whereas specializing in Trump’s character traits when requested to mocked him.
All of its replies can be found on our GitHub repository.
Total, the bias exists however stays much less pronounced than the pro-U.S. slant in Claude/ChatGPT, or the pro-China positioning in DeepSeek/Qwen, for instance. Builders, after all, will be capable to finetune this mannequin so as to add as a lot censorship, freedom or bias as they need—because it occurred with DeepSeek-R1, which was finetuned by Perplexity AI to offer a extra pro-U.S. bias on its responses.
Agentic work and net searching
MiniMax-M1’s net searching capabilities are a very good function for these utilizing it by way of the official chatbot interface. Nonetheless, they can’t be mixed with the considering capabilities, severely hindering its potential.
When tasked with making a two-week Venezuela journey plan on a $3,000 funds, the mannequin methodically evaluated choices, optimized transportation prices, chosen applicable lodging, and delivered a complete itinerary. Nonetheless, the prices, which have to be up to date in actual time, weren’t based mostly on actual info.
Claude produces higher-quality outcomes, but it surely additionally expenses for the privilege.
For extra devoted duties, MiniMax gives a devoted brokers tab with capabilities akin to Manus—performance that ChatGPT and Claude have not matched. The platform offers 1,000 free AI credit for testing these brokers, although that is simply sufficient for gentle testing duties.
We tried to create a customized agent for enhanced journey planning—which might have solved the issue of the dearth of net looking capabilities within the final immediate—however exhausted our credit earlier than completion. The agent system exhibits large potential, however requires paid credit for critical use.
Non-mathematical reasoning
The mannequin reveals a peculiar tendency to over-reason, typically to its personal detriment. One check confirmed it arriving on the appropriate reply, then speaking itself out of it via extreme verification and hypothetical situations.
We prompted the same old thriller story from the BIG-bench dataset that we usually use, and the ending consequence was incorrect because of the mannequin overthinking the difficulty, evaluating potentialities that weren’t even talked about within the story. The entire Chain of Thought took the mannequin over 700 seconds—a report for this type of “easy” reply.
This exhaustive method is not inherently flawed, however creates prolonged wait occasions as customers watch the mannequin work via its chain of thought. As a thumbs-up function, not like ChatGPT and Claude, MiniMax shows its reasoning course of transparently—following DeepSeek’s method. The transparency aids debugging and high quality management, permitting customers to determine the place logic went astray.
The issue, together with MiniMax’s entire thought course of and reply can be found in our GitHub repo.
Verdict
MiniMax-M1 isn’t excellent, but it surely delivers fairly good capabilities for a free mannequin, providing real competitors to paid companies like Claude in particular domains. Coders will discover a succesful assistant that rivals premium choices, whereas these needing long-context processing or web-enabled brokers acquire entry to options sometimes locked behind paywalls.
Inventive writers ought to look elsewhere—the mannequin produces useful however uninspired prose. The open-source nature guarantees important downstream advantages as builders create customized variations, modifications, and cost-effective deployments unimaginable with closed platforms like ChatGPT or Claude.
This can be a mannequin that may higher serve customers requiring reasoning duties—however remains to be an excellent free various for these searching for a chatbot for on a regular basis use that’s not actually mainstream.
You may obtain the open supply mannequin right here.
