the $10b shortcut
we are currently witnessing a structural bifurcation in the global ai landscape. on one side, you have the closed-source giants (openai, anthropic, google) scaling dense models behind opaque apis. on the other, the open-source ecosystem has been playing catch-up, mostly through brute-force scaling or finetuning llama derivatives.
but every so often, a model drops that doesn't iterate but rather it disrupts.
the release of minimax m2 is one of those moments. it poses a very uncomfortable question for the proprietary labs, asking how a model that ranks in the top five of the artificial analysis benchmark, territory usually reserved for gpt-4 and claude opus, can exist with open weights? [1]
i have been testing m2 extensively. it is currently the third most-used model on openrouter and let me be clear, this isn't hype. it is a pragmatic recognition that we have finally hit the "intelligence-cost quadrant" sweet spot. we have been looking for pareto optimality where high reasoning capabilities intersect with dirt-cheap inference. m2 occupies this space decisively.
here is my deep dive into why this 230-billion parameter monster (that only thinks it is a 10-billion parameter model) is the future of agentic intelligence.
---
1. the "mini/max" paradox
the nomenclature "minimax" is the governing engineering principle behind m2.
- "mini" refers to the active computational footprint.
- "max" refers to the knowledge capacity.
the architecture relies on a massive mixture-of-experts (moe) design. the total parameter count is a staggering 230 billion. in a traditional dense model (like llama 3 70b), every single one of those parameters would fire for every token generated. that is computationally expensive and slow.
m2, however, uses an aggressive sparsity ratio of approximately 23:1. it only activates 10 billion parameters per token generation step [3].
the math of efficiency
to visualise why this matters, consider the cost of inference (C). in a dense model, cost scales with total parameters (P_{total}). in m2, it scales with active parameters (P_{active}).
$C_{M2} \propto P_{active} \approx \frac{P_{total}}{23}$
this simple equation is why m2 changes the economics of ai. it holds the encyclopedic breadth of a 200b+ model (the "max") but runs with the latency and cost profile of a small 10b model (the "mini").
for autonomous agents, this is critical. agents require loops. they plan, they act, they observe, they refine. this cycle can consume thousands of tokens per task. if your underlying model is claude 3 opus, you go broke. if it is llama 3 8b, the agent is too stupid to finish the task. m2 hits the sweet spot because it is smart enough to handle long-horizon tasks but roughly 92% cheaper than comparable proprietary models [7].
---
2. the great attention pivot and why m2 dropped "lightning"
here is a nuanced detail that many people miss. the predecessor, minimax m1, was famous for using "lightning attention", a linear attention mechanism that allowed for massive 1m+ token contexts.
m2 dropped it.
m2 reverted to standard full attention (softmax). why? because in the real world of coding and "needle-in-a-haystack" retrieval, precision beats theoretical efficiency.
linear attention approximates the interaction between tokens to achieve O(n) complexity. full attention has quadratic complexity O(n^2), meaning it's computationally heavier.
$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$
see that QK^T term? that's the model comparing every token to every other token. minimax engineers realised that for an agent to correctly identify a single bug in 50 files of code, you cannot afford the "lossy" compression of linear attention [4]. you need the brutal exactness of the n \times n matrix. this decision to prioritise accuracy over raw speed is why m2 feels so much sharper at coding tasks than m1 did.
---
3. the death of chain-of-thought and the rise of "interleaved reasoning"
we need to talk about how this thing thinks. since mid-2024, the industry has been obsessed with chain-of-thought (cot), the idea that the model should vomit out a "reasoning preamble" before giving you an answer.
the problem with standard cot is that, it is linear.
input → giant block of thinking → final output.
this is fine for solving a math riddle, but it is terrible for agents. agents need to interact with the world.
minimax m2 introduces interleaved reasoning. it doesn't separate thinking and doing; it blends them into a continuous loop, effectively modeling the task as a markov decision process (mdp):
input → think → tool call (act) → observation → think → tool call (act) → answer.
this is a subtle but profound shift. in this paradigm, "thinking" isn't a preamble; it is a pervasive thread. if the agent tries to scrape a website and gets a 403 forbidden error, it doesn't fail. it sees the observation, "thinks" about it ("i need to try a different endpoint") and acts again.
in my testing, this state preservation is the killer feature. the reasoning traces are kept as part of the conversation history object. at step 50 of a complex task, m2 can look back at step 10 and see why it made a decision, not just what it did [1].
---
4. the stability algorithm behind cispo
why is m2 so stable during these long loops? it comes down to how it was trained.
training reasoning models with reinforcement learning (rl) is notoriously unstable. as models generate longer sequences of thought, the variance in gradients explodes. the standard fix is to use algorithms like ppo, which use "clipping" to stop the model from changing too much, too fast.
the problem? ppo clips the update. this often suppresses those "aha!" moments where the model makes a massive leap in logic because the algorithm interprets that leap as instability [8].
minimax uses a novel algorithm called cispo (clipped importance sampling policy optimisation).
$L_{CISPO}(\theta) = \mathbb{E}\left[\min\left(r_t(\theta)A_t, \text{clip}(r_t(\theta), 1-\epsilon, 1+\epsilon)A_t\right)\right]$
without getting bogged down in the full derivation, the "magic" is that cispo clips the weight but detaches it from the gradient. it tells the model: "this was a huge, high-variance move. i will let you learn from it, but i won't let you blow up the training run over it."
meta ai recently released a paper on scaling rl compute and explicitly identified cispo as the "most stable way to scale reinforcement learning steps" [5]. when meta validates your math, you know you are on to something.
---
5. benchmarks and the "tau squared" coding dominance
the report request mentioned the "tau squared" benchmark. this refers to tau-bench (\tau^2), the current gold standard for agentic reliability.
- tau-bench: m2 scores 87% in the telecom domain. this measures reliability in multi-turn tool use. most models lose track of the user's constraints after turn 3. m2 holds the state.
- swe-bench verified: this is the big one for autonomous coding. m2 scores 69.4%. for context, claude 3.5 sonnet is around 77% and gpt-4o is around 75%. m2 is within striking distance of the absolute state-of-the-art, while being open-weights and significantly cheaper to run [6].
- browsecomp: in web navigation agents, m2 scored 44.0%, which ironically nearly triples the score of claude 3.5 sonnet (19.6%) on that specific evaluation.
---
to wrap up
by decoupling total knowledge from active compute (230b/10b), minimax has solved the economic blocker for autonomous agents. by implementing interleaved reasoning, they have solved the cognitive blocker. and by using cispo, they have turned the "black magic" of rl training into a rigorous science.
for developers, the choice is becoming clearer. you can pay the "proprietary tax" for the absolute peak of intelligence or you can use m2 to get 95% of the performance at a fraction of the cost, with full control over your infrastructure.
in the intelligence vs. price quadrant, m2 stands alone. it is the best bang for the buck we have ever seen.
---
references
[1] minimax technical report, "the m2 architecture and performance," 2024. [2] artificial analysis, "ai model leaderboard q4 2024." [3] openrouter documentation, "minimax m2 model card and pricing," 2024. [4] "why m2 returned to full attention," minimax engineering blog, 2024. [5] meta ai research, "the art of scaling rl compute," arxiv, 2024. [6] swe-bench leaderboard, "verified split results," 2024. [7] user comparisons on "the intelligence-price quadrant," reddit/localllama, 2024. [8] "stabilizing rlhf with cispo," minimax research, 2024.