MiniMax M3 vs Kimi K2.7 Code vs Nemotron 3 Ultra: The Best Open-Weight AI Models of 2026 Compared

Three massive open-weight models dropped within two weeks of each other. One from a Shanghai startup nobody outside of AI circles knew a year ago. One from Beijing’s most watched AI lab. One from the company that makes the chips everyone else is running their models on. Each is making a different bet on what matters most in AI agents right now — and all three are worth your serious attention.

---

---

Why This Comparison Matters Right Now

If you’ve been tracking the open-weight AI space for the past six months, you already know the story is moving faster than anyone predicted. The same capabilities that required a $15/M-token Claude Opus subscription twelve months ago are now available as downloadable weights you can run on your own servers.

But June 2026 was something else entirely. In the space of twelve days, three major open-weight releases landed within days of each other: MiniMax M3 on June 1st, Nemotron 3 Ultra on June 4th, and — literally yesterday — Kimi K2.7 Code on June 12th. Combined, they represent arguably the most significant two-week stretch in open-source AI history, and each one is making a different architectural and philosophical bet on what the next generation of AI agents needs to do.

This isn’t just a benchmark roundup. We’re going to look at how these models actually behave in production: how they fare on the agent frameworks developers are genuinely using — OpenClaw, Hermes Agent, and the increasingly popular VIBE Coding workflow — and where each model will save you money or time versus where it’ll quietly let you down.

Let’s start with who these models actually are.

---

The Contenders at a Glance

Before diving deep, here’s the 30-second summary of what each model is and why it exists:

MiniMax M3 is Shanghai-based MiniMax’s play to be the first open-weight model that genuinely combines three things at once: frontier-level coding performance, a one-million-token context window, and native multimodality — including image, video, and desktop computer control. It ships under an open-weights license with subscription pricing starting at $20/month.

Kimi K2.7 Code is Moonshot AI’s (Beijing) fifth major release in under a year — a laser-focused coding upgrade to the already-impressive K2.6, cutting reasoning token usage by approximately 30% while pushing key agent benchmarks meaningfully higher. It dropped on Hugging Face yesterday under a Modified MIT license, priced at $0.95/$4.00 per million tokens.

Nemotron 3 Ultra is NVIDIA’s answer to the question: what if the chip-maker built the model too? At 550 billion parameters, it’s the most capable open-weight model to come out of a US-based lab, scoring 47.7 on the Artificial Analysis Intelligence Index — a score that puts it in the same tier as Claude Opus 4.6 and Kimi K2.6. It was released June 4th under the Linux Foundation’s OpenMDW-1.1 license.

Three models. Three very different origin stories. All worth running.

---

MiniMax M3

Who Made It and Why

MiniMax isn’t a household name for most people outside of AI circles, but inside them, the Shanghai-based company has been increasingly hard to ignore. They listed on the Hong Kong Stock Exchange in January 2026, and they’ve been building toward M3 for the better part of a year. The pitch is ambitious to the point of being almost aggressive: M3 is positioned as the first open-weight model to combine frontier coding, a million-token context window, and native multimodality in a single system — and they launched it at a price point that makes closed-model pricing look expensive.

The Architecture: MiniMax Sparse Attention

The headline technical story here is the MiniMax Sparse Attention (MSA) architecture. Standard transformer attention is quadratic — every token attends to every other token, which means doubling your context roughly quadruples your compute bill. At a million tokens, that math becomes brutal.

MiniMax built M3 on its new MSA design, which cuts per-token compute at 1M context to approximately one-twentieth of the prior generation, with more than 9× faster prefill and more than 15× faster decoding. That’s not a marginal improvement. That’s the difference between a million-token context window being a theoretical spec and it being something you can actually ship to production without your GPU bill going vertical.

The model is built on MiniMax Sparse Attention (MSA), which replaces full attention with KV-block selection to cut per-token compute at long context, with substantially faster prefill and decode while retaining quality across most tasks.

Interestingly, MiniMax killed sparse attention in their M2 generation and brought it back specifically for M3 — suggesting they had the architecture working but needed more training data or post-training alignment work to make it competitive. M3 suggests they figured it out.

What Makes M3 Actually Different

MiniMax released M3 pairing frontier-tier coding and agentic performance with a 1-million-token context window and native multimodality for a fraction of the cost of leading proprietary models.

The multimodality angle is worth pausing on. M3 doesn’t just accept images — it ingests images, video, and can operate a desktop computer natively. For vibe coding workflows where you’re passing in a screenshot of a UI and saying “build me this,” that’s genuinely useful in a way that a text-only coding model isn’t.

MiniMax-M3 is MiniMax’s frontier multimodal coding and agentic model, built on the MSA architecture. It supports up to a 1M-token context window and accepts image and video inputs. The model is designed for code generation, agentic workflows, tool use, long-context understanding, and multi-step reasoning.

The VIBE Benchmark: MiniMax’s Self-Created Standard

One thing worth flagging: MiniMax didn’t just release a model — they released a new benchmark. The VIBE (Visual & Interactive Benchmark for Execution) framework is MiniMax’s answer to what they see as a gap in existing evaluation frameworks. Unlike traditional benchmarks like SWE-bench and Terminal-bench, which focus on static code correctness or command-line–level task completion, VIBE automatically evaluates the interaction logic and visual presentation of generated applications in a real execution environment, providing a more faithful assessment of real user experience.

It’s a self-serving benchmark, sure — but the underlying critique isn’t wrong. SWE-bench tells you whether a model can resolve GitHub issues; it doesn’t tell you whether a model can build a running, visually coherent web app from a screenshot. M3 scores well here, which makes sense given its native multimodal architecture.

MiniMax M3 Benchmarks

MiniMax M3 scores 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1, 34.8% on SWE-fficiency, and 83.5 on BrowseComp. The BrowseComp number in particular is striking — 83.5 puts it ahead of GPT-5.5 on autonomous web browsing tasks, which is exactly the kind of capability you want in an agent that’s doing research-heavy coding work.

MiniMax M3 scores 59.0% on SWE-Bench Pro, beating GPT-5.5 and Gemini 3.1 Pro and approaching Claude Opus 4.7. It also scores 66.0% on Terminal Bench 2.1, 34.8% on SWE-fficiency, 28.8% on KernelBench Hard, and 74.2% on MCP Atlas.

A caveat worth mentioning: MiniMax’s comparison baseline in its own materials uses Claude Opus 4.7, not the more recently released Opus 4.8. That framing is not inaccurate, but developers evaluating M3 against the current benchmark ceiling should use Opus 4.8 figures, which place M3 further from the frontier than the launch announcement implies.

Still — 59% on SWE-Bench Pro from an open-weight model with a million-token context is remarkable, and the independent data largely confirms the company-reported numbers are in the right ballpark.

Pricing

At launch, MiniMax M3 was listed on OpenRouter at $0.60 per million input tokens and $2.40 per million output tokens, with a temporary 50% promotional discount bringing it to roughly $0.30 input and $1.20 output per million tokens — a fraction of frontier closed models like Claude Opus and GPT-5.5.

For context: Claude Opus 4.8 runs $5/$25 per million tokens. M3 at promotional pricing is roughly 1/20th of that cost. Even at standard pricing, you’re looking at 1/10th. The cost story is as compelling as the capability story.

---

Kimi K2.7 Code

Who Made It and Why

Moonshot AI launched yesterday — June 12, 2026 — with what is the fifth major release in under a year for the Beijing-based company, and they’ve positioned their models around three pillars: agentic capabilities, extended context handling, and multimodal inputs. The Kimi K2 family has become one of the most-watched open-weight lineages of 2026, and K2.7 Code is its sharpest iteration yet.

The name change is meaningful. This is the first time Moonshot put “Code” explicitly in the model name. They’re not pretending K2.7 is a general-purpose model — it’s tuned for engineering, not broad chat, and they want you to know it.

Architecture: The Same Trillion-Parameter Foundation, Tuned Harder

Kimi K2.7 Code is a 1-trillion-parameter Mixture-of-Experts model with 32B active parameters per token and 384 experts, with a 262,144-token context window, inherited from K2.6, with automatic context compression for sustained long-horizon sessions.

The architecture itself hasn’t changed dramatically from K2.6 — it’s still the 1T MoE framework that made Kimi K2.6 the highest-ranked open-weight model on the Artificial Analysis Intelligence Index earlier this year. What K2.7 Code represents is a targeted refinement: the same chassis, with the engine retuned specifically for agentic coding workflows.

Kimi K2.7 Code is Moonshot AI’s coding-focused agentic model built on Kimi K2.6. It improves real-world long-horizon coding task completion, instruction following, and token efficiency while reducing thinking-token usage by approximately 30% versus Kimi K2.6.

That 30% reduction in thinking tokens isn’t a small deal. In a long-running coding agent session where the model is spinning through hundreds of turns, fewer thinking tokens means lower cost and faster wall-clock time. If you’re running something like a full-repository refactor overnight, K2.7’s efficiency gains over K2.6 compound significantly.

What K2.7 Actually Improves

Moonshot’s announcement leads with three numbers: +21.8% over K2.6 on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite, alongside a claim of roughly 30% lower reasoning-token usage versus K2.6.

These are first-party numbers, run on Moonshot’s own benchmark suites. The honest note here is that as of June 13, 2026 — one day after release — there are no independent third-party numbers for K2.7 on the standard public suites — SWE-bench Verified, SWE-bench Pro, Terminal-Bench, LiveCodeBench, GPQA Diamond, AIME, or MMLU-Pro. That verification will come in the days ahead. Based on K2.6’s trajectory (which did produce competitive independent benchmarks), the internal numbers are likely directionally accurate.

The standout independent-adjacent result so far: K2.7-Code scored 81.1 on MCP Mark Verified, beating Claude Opus 4.8’s 76.4%. That suite tests correct tool invocation through the Model Context Protocol — CI checks, ticket updates, and file edits in one loop. The fact that an open-weight model is now beating Opus 4.8 on MCP tool use is, by any reasonable measure, a watershed moment for open-source AI.

The Kimi Code Platform Story

One element that gets underreported in model comparisons: K2.7-Code launches inside Kimi Code — Moonshot’s open-source terminal agent — with membership plans listed from $19/month. Moonshot is explicitly competing on the full stack here: model, CLI, and subscription economics. This is Cursor-vs-Kimi economics, and it matters for teams evaluating their AI coding workflow holistically rather than just the raw model.

The Kimi Code CLI itself is worth trying. It’s shell-aware, supports MCP server integration, and the open-source licensing means you’re not locked into Moonshot’s API if you want to self-host.

Multimodal Capabilities

Like M3, K2.7 Code isn’t limited to text. Developers can upload screenshots, diagrams, product mockups, or even videos and ask the model to generate code based on them. This makes it useful for frontend development, debugging visual issues, and reverse-engineering interfaces. The vision capability is real and genuinely useful for UI-centric coding tasks, though K2.7’s multimodal story is slightly narrower than M3’s (which includes desktop computer use).

Pricing

Pricing for Kimi K2.7 Code is $0.95 per million input tokens, $4.00 per million output tokens, and $0.19 per million on cache hits, on the Moonshot API. Free weights are available on Hugging Face for self-hosting.

For a trillion-parameter model, $0.95 input is very competitive. The output price ($4.00) is higher than M3 in absolute terms, but K2.7’s 30% reduction in reasoning tokens means real-world cost per completed task can be lower than the raw per-token number suggests. Pricing should be modeled against your specific workload, not just the headline rate.

---

Nemotron 3 Ultra

Who Made It and Why

Here’s the thing about Nemotron 3 Ultra that most coverage underplays: this is NVIDIA building a frontier model. Not a chip company dabbling in AI, not a research lab attached to an infrastructure business — NVIDIA, whose revenue depends on everyone else’s AI training runs, decided to build and release a competitive frontier model under a fully open license. That’s a meaningful statement about where the market is going.

On June 4, 2026, NVIDIA released Nemotron 3 Ultra, a fully open 550 billion parameter reasoning model built specifically for long running agents.

On June 4, 2026, NVIDIA quietly dropped Nemotron 3 Ultra to Hugging Face, two days after Jensen Huang announced it from the Computex stage in Taipei. 550 billion parameters. 55 billion active per forward pass. Over 300 tokens per second. The highest Intelligence Index score of any US-developed open-weight model, ever.

Architecture: The Hybrid Mamba-Transformer That Changes Everything

This is where Nemotron 3 Ultra gets technically interesting. While M3 uses sparse attention to handle long contexts and K2.7 stays on the proven MoE transformer path, Nemotron Ultra takes a different route entirely.

NVIDIA released Nemotron 3 Ultra as a 550B-parameter Mixture-of-Experts model with 55B active parameters, optimized for orchestrating complex, long-running agent workflows. Architectural innovations include hybrid Mamba-Transformer layers for efficient long-context handling, NVFP4 quantization for cross-architecture GPU deployment with up to 5x higher throughput, LatentMoE for expert routing, and multi-token prediction for improved generative speed in multi-turn tasks.

The hybrid Mamba-Transformer approach is significant. Mamba layers handle sequential dependencies more efficiently than standard attention at long contexts — they scale linearly rather than quadratically with sequence length. Combining that with transformer attention for tasks that benefit from it, and wrapping the whole thing in a MoE architecture, lets NVIDIA hit an unusual combination: high intelligence, high throughput, and relatively low per-token cost.

The NVFP4 quantization is another differentiator. On Blackwell GPUs (H100 successors), it runs with native FP4 math. On Hopper (H100), it falls back to W4A16. Either way, Nemotron 3 Ultra achieves 5.9x higher inference throughput compared to GLM-5.1, 4.8x faster than Kimi K2.6, and 1.6x faster than Qwen-3.5 on 8K input / 64K output token settings, while attaining on-par accuracy across a wide range of agentic and reasoning benchmarks.

Read that again: 4.8x faster than Kimi K2.6 at comparable accuracy. For teams running production agents, that throughput differential translates directly to per-task cost.

Benchmarks and Intelligence Index

Nemotron 3 Ultra scores 47.7 on the Artificial Analysis Intelligence Index — well ahead of the next strongest US open weights models, Gemma 4 31B at 39.2, Nemotron 3 Super at 36.0, and gpt-oss-120b at 33.3.

The honest caveat is right there in the same paragraph: it’s still behind the Chinese-led open-weights frontier (Kimi K2.6 at 53.9). This is a US-first story, and it’s a good one — but Nemotron Ultra isn’t the best open-weight model in the world right now. It’s the best American one, and by a wide margin.

On agentic tasks, Nemotron 3 Ultra posts 90.0 on PinchBench and 56.0 on ProfBench Search. NVIDIA team reserved both as held-out generalization gates, scored only once on the final model. It scores 71.9 on SWE-Bench Verified and 56.4 on Terminal Bench 2.1. On reasoning, it scores 570.0 on IOI 2025, which NVIDIA frames as top-3-human-level competitive programming.

The SWE-Bench Verified score of 71.9% is particularly competitive. That trails Claude Fable 5 and GPT-5.5 at the closed-model frontier, but leads most open-weight alternatives — and importantly, it’s measured consistently across multiple agent frameworks.

On AA-Omniscience, it records the highest non-hallucination score in the set at 78.7, suggesting a lower tendency to answer when uncertain. Long context holds up at scale — the model scores 94.7 on RULER at 1 million tokens.

That hallucination figure is worth highlighting. In production agent workflows, a model that knows when it doesn’t know something is often more valuable than one that scores a few percentage points higher on benchmarks but confidently hallucinates its way through edge cases.

The Open Training Data Story

NVIDIA has released — cumulatively across the three Nemotron 3 launches — 50 million supervised fine-tuning samples, 2 million reinforcement learning tasks, and 55 RL environments. That level of openness is unusual for a frontier-class model family, and it’s the thing the AI research community has responded to most strongly.

If you want to fine-tune or extend Nemotron Ultra, you’re not doing it in the dark. The training recipes, data, and evaluation environments are all public. For enterprise teams that need domain-specific performance and have the capability to fine-tune, this is a meaningful advantage that neither M3 nor K2.7 Code currently matches.

Pricing

DeepInfra has a pre-release endpoint already active, with pricing at $0.37/M input and $1.08/M output — better than median for this size tier. OpenRouter has it indexed and accessible. Enterprise integration is available via NVIDIA NIM microservices at build.nvidia.com.

Nemotron 3 Ultra runs at $0.50 per million input tokens and $2.50 output on some providers, offering strong price-performance for a model at this intelligence level.

---

Head-to-Head Benchmarks

Here’s a side-by-side look at the core numbers across the three models, plus context on what each benchmark actually measures:

SWE-Bench Pro (Real GitHub Issue Resolution)

SWE-Bench Pro tests the ability to resolve actual GitHub issues filed after a model’s training cutoff — reducing data contamination risk compared to earlier SWE-Bench variants. It’s the closest thing to “can this model fix real bugs in real codebases?” that the benchmark community has produced so far.

• MiniMax M3: 59.0% (company-reported)
• Nemotron 3 Ultra: ~56-71.9% (SWE-Bench Verified); SWE-Bench Pro numbers pending independent confirmation
• Kimi K2.7 Code: K2.6 posted 58.6%; K2.7’s internal gains suggest similar or higher — independent score pending

Terminal-Bench 2.1 (Multi-Step CLI Tasks)

Terminal-Bench tests what agents actually do in production: multi-step shell tasks in live terminal environments. This is closer to “can it run a CI pipeline” than “can it autocomplete code.”

• MiniMax M3: 66.0%
• Nemotron 3 Ultra: 56.4%
• Kimi K2.7 Code: K2.6 led at 67.2%; K2.7 expected to improve on this

On Terminal-Bench, M3 and K2.7 are in a similar tier, both ahead of Nemotron Ultra. This makes sense — M3 and the Kimi family have been optimized specifically for long-horizon coding tasks, while Nemotron Ultra’s design priorities are broader.

MCP Tool Use (Model Context Protocol)

MCP tool use is increasingly the benchmark that matters most for real agent deployments. Correct tool invocation isn’t just about benchmark scores — it’s about whether your agent actually executes the right actions without hallucinating function names or parameters.

• Kimi K2.7 Code: 81.1 on MCP Mark Verified (beats Claude Opus 4.8’s 76.4)
• MiniMax M3: 74.2 on MCP Atlas (company-reported)
• Nemotron 3 Ultra: Strong BFCL V4 scores (function calling), exact MCP Mark numbers not yet published

K2.7 wins this category clearly. Moonshot’s agentic fine-tuning has specifically targeted tool calling patterns, and it shows.

Artificial Analysis Intelligence Index (Composite Score)

This 10-evaluation composite covers reasoning, knowledge, mathematics, and coding — essentially a weighted average of a model’s general intelligence.

• Kimi K2.6 / K2.7 family: 53-54 (K2.7 data pending)
• Nemotron 3 Ultra: 47.7
• MiniMax M3: Not yet rated (BenchLM places M3 at 76/100 on provisional leaderboard, #29 of 122)

Speed (Tokens per Second)

This is where Nemotron Ultra’s architecture tells a different story:

• Nemotron 3 Ultra: 300+ tokens/second on GB200, 5.9x faster than GLM-5.1, 4.8x faster than Kimi K2.6
• MiniMax M3: ~100 tokens/second at 1M context
• Kimi K2.7 Code: Comparable to K2.6 (throughput improvements from token efficiency rather than raw speed)

For high-throughput production deployments where you’re running many agents in parallel, Nemotron Ultra’s throughput advantage is substantial.

---

Real-World Agent Performance: Hermes, OpenClaw & WildClawBench

Benchmarks are one thing. How these models actually behave in the agent frameworks developers are deploying today is another.

OpenClaw: The Agent Framework Taking Over

If you haven’t heard of OpenClaw yet, you will. In just two months, OpenClaw garnered 247,000 GitHub Stars, becoming an AI agent platform eagerly adopted by companies in Silicon Valley and China. It’s local-execution, model-agnostic, and integrates with messaging apps — three characteristics that make it meaningfully different from SaaS-based AI assistants.

OpenClaw is model-agnostic by design, which means the quality of your experience depends almost entirely on which LLM you drop into the back end. This is where model choice becomes directly consequential.

The PinchBench OpenClaw benchmark runs 23 tasks across code execution, content creation, research, and system tools. The test includes 23 tasks covering code execution, content creation, and system tools — open-source and reproducible, using data from PinchBench’s OpenClaw agent tests.

On PinchBench’s OpenClaw evaluation:

• Nemotron 3 Ultra: 90% on PinchBench Agent Productivity (ties Kimi K2.6 — the best Chinese open model on task completion according to the benchmark)
• Kimi K2.7 Code: Expected to be competitive with or exceed K2.6’s strong showing
• MiniMax M3: Strong agentic scores generally; the 5x cost advantage over Kimi K2.6 in Composio’s real-world tool tests is worth noting

A Composio real-world comparison of M3 vs K2.6 (the K2.7 predecessor) found something interesting: M3 cost $0.81 across 25 Composio tasks, while Kimi cost $4.08 — about 5x more. M3 had the clearer edge on hard terminal coding; everyday SaaS tool orchestration was effectively even.

That cost gap isn’t just an economics story — it means you can run 5x more agent iterations for the same budget, which matters for exploration-heavy coding tasks where iteration speed is the bottleneck.

Hermes Agent: The Framework That Doesn’t Play Favorites

WildClawBench, published in May 2026 by InternLM, is one of the more rigorous independent agent evaluations available right now. The benchmark tests what actually matters: can an AI agent do real work, end-to-end, without hand-holding? It runs the same 60-task suite under four different agent harnesses — OpenClaw, Claude Code, Codex CLI, and Hermes Agent — separating model capability from harness scaffolding.

The Hermes harness specifically is interesting because it was designed to test models independently of any vendor’s agent scaffolding. It’s a clean signal on the underlying model quality.

Nemotron 3 Ultra achieves SWEBench Verified scores between 65% and 70.4% across Pi, OpenHands, Hermes, OpenCode, and Mini SWE Agent — consistent performance regardless of which framework you deploy.

That consistency across harnesses is a meaningful quality signal. Many models that score well in their own CLI degrade significantly when you drop them into a different agent framework. Nemotron Ultra’s architecture appears to be robust to framework changes — likely because NVIDIA deliberately trained across multiple agent harnesses per task type rather than optimizing for one.

For M3 and K2.7 Code, WildClawBench-specific numbers are still filtering through the community, but the K2 family’s historical strength in agentic evaluations and K2.7’s MCP improvements suggest strong Hermes harness results.

The Framework Conclusion

If you’re building on OpenClaw and cost matters: M3 is your model. If you’re doing MCP-heavy pipelines: K2.7 Code is the leader. If you need framework-agnostic consistency across Hermes, OpenClaw, and others simultaneously: Nemotron Ultra’s harness-independent consistency is the right call.

---

Vibe Coding: Which Model Actually Builds Apps Well?

“Vibe coding” has become the shorthand for a style of AI-assisted development where you’re building entire features or small applications iteratively from natural language prompts, often starting from screenshots, mockups, or rough descriptions. It’s less about precise code generation and more about the model’s ability to hold a coherent product vision across many turns while producing working, visually coherent output.

Each of these three models approaches vibe coding differently, and the differences matter.

MiniMax M3: Strongest Multimodal Foundation

M3 is purpose-built for the visual dimension of vibe coding. MiniMax introduced the VIBE (Visual & Interactive Benchmark for Execution) specifically to measure a model’s ability to build complete, runnable applications from zero to one — automatically evaluating the interaction logic and visual presentation of generated applications in a real execution environment.

That M3 introduced this benchmark is telling. MiniMax clearly believes that standard coding benchmarks — which test whether code passes unit tests — miss the question that actually matters for vibe coding: does the app work and look right?

In practice, M3’s native video/image input means you can paste a Figma screenshot or a recorded user session and ask the model to build from it directly. That’s a fundamentally different workflow than describing what you want in words. For frontend-heavy development — React components, web app UIs, mobile-first layouts — M3’s visual input capability changes the loop in ways that pure-text models can’t match.

The 1M token context window also matters for vibe coding in a specific way: you can fit an entire medium-sized codebase in context, meaning the model can refactor across files without losing coherence about what it’s building. This is one of the real failure modes in vibe coding with smaller-context models — the code starts to diverge from itself across turns as the model loses sight of earlier architectural decisions.

Kimi K2.7 Code: Best Tool-Integrated Vibe Coding

Developers can upload screenshots, diagrams, product mockups, or even videos and ask Kimi K2.7 to generate code based on them — useful for frontend development, debugging visual issues, and reverse-engineering interfaces.

K2.7’s advantage in vibe coding is its MCP tool integration. When you’re building through Kimi Code CLI, the model can loop through a real terminal: running the app, checking the output, reading error logs, and iterating — all within the same session. That end-to-end loop with real execution feedback is what separates “generates code” from “actually builds the thing.”

The 30% token reduction also matters for vibe coding specifically because vibe coding sessions tend to be long and conversational. A model that uses fewer tokens per reasoning step can sustain longer sessions at lower cost, which is directly aligned with how iterative app building actually works.

Nemotron 3 Ultra: Powerful but Less Specialized

Nemotron Ultra is a strong all-rounder and handles multimodal inputs well, but its design priorities were optimization for agent consistency and reasoning depth rather than the visual interface of vibe coding. For pure vibe coding workflows — screenshot in, working app out — M3 is the more natural fit.

Where Nemotron Ultra excels in development workflows is in the architectural and reasoning-heavy phases of building: designing system components, debugging complex state management issues, or working through a tricky algorithm. It’s less “build this UI from this screenshot” and more “help me design the architecture and reason through the edge cases.”

Vibe Coding Verdict

For vibe coding, the ranking goes: M3 (best visual integration, largest context, cheapest) → K2.7 Code (best real-execution feedback loop, strongest tool use) → Nemotron Ultra (excellent reasoning assistant, less specialized for visual app building).

If you’re doing vibe coding primarily in a visual-heavy stack (React, Vue, mobile), M3 is your model. If you’re building backend-heavy apps or systems where the iteration loop involves running and testing the actual code, K2.7 Code’s MCP integration is the edge.

---

Pricing & Cost-to-Performance Analysis

Let’s look at the actual cost math, because the headline benchmarks only make sense in the context of what you’re paying.

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context	License
MiniMax M3 (promo)	$0.30	$1.20	1M	Open weights
MiniMax M3 (standard)	$0.60	$2.40	1M	Open weights
Kimi K2.7 Code	$0.95	$4.00	256K	Modified MIT
Nemotron 3 Ultra (DeepInfra)	$0.37	$1.08	1M	OpenMDW-1.1
Claude Opus 4.8	$5.00	$25.00	—	Proprietary
GPT-5.5	~$10.00+	~$30.00+	—	Proprietary

The cost story becomes more nuanced when you account for efficiency. K2.7’s 30% token reduction means a session that uses 1M output tokens on K2.6 uses ~700K on K2.7. At $4.00/M output, that’s $4.00 vs $2.80 — not just a rate difference but an efficiency savings. Over long agent runs, that compounds significantly.

Nemotron Ultra at $0.37/$1.08 from DeepInfra is remarkably affordable for a 550B model with 47.7 intelligence index — partly a function of its throughput advantage (5.9x faster means lower cost per completed task even if per-token rates were the same).

M3 on promotional pricing is the cheapest at absolute rates, though standard pricing at $0.60/$2.40 is still excellent value for a million-token context window with frontier-tier coding.

For most production deployments, the real cost calculation isn’t tokens-per-dollar but completions-per-dollar. Build a small representative test suite of your actual tasks and price each model against that. The headline rates are a starting point, not the answer.

---

Who Should Use Which Model?

The three models aren’t really competitors for the same use case. Here’s a pragmatic breakdown:

Choose MiniMax M3 if:

• You need the largest context window for large codebases, long documents, or multi-file projects
• Your workflow is visually driven (screenshots → code, mockups → implementation)
• Cost is your primary constraint and you need to run many iterations
• You’re building browser agents or autonomous research pipelines that benefit from BrowseComp’s strength
• You want a single model that handles text, images, and video without switching

Choose Kimi K2.7 Code if:

• You’re building MCP-integrated pipelines and need the best tool invocation reliability
• You’re doing long-horizon agentic coding with multi-step terminal workflows
• You’re deploying through Kimi Code CLI and want the tightest model-to-harness integration
• You want open-weight model flexibility with Modified MIT licensing
• Token efficiency over long sessions matters (the 30% reduction compounds across large agentic runs)

Choose Nemotron 3 Ultra if:

• You’re in a US-based enterprise and geopolitical data concerns rule out Chinese-origin models
• You need model fine-tuning with full access to training data, SFT samples, and RL environments
• Framework consistency matters — you’re deploying across Hermes, OpenClaw, and other harnesses simultaneously
• Throughput is critical for high-volume concurrent agent deployments
• You’re doing reasoning-heavy work: competitive programming, complex architecture, graduate-level reasoning tasks
• You want the best open-weight American model, full stop

The honest edge case: If you’re in a small startup or indie developer context with flexible data requirements and cost is your primary constraint, M3 at promotional pricing is genuinely hard to beat. If you’re in a regulated US enterprise context, Nemotron Ultra is the safe default, and it’s competitive enough that you’re not sacrificing much by avoiding the Chinese-origin models.

---

The Elephant in the Room: Data Privacy

This comparison would be incomplete without addressing a topic that’s increasingly relevant for enterprise deployments.

MiniMax is headquartered in Shanghai. Under China’s National Intelligence Law enacted in 2017, every Chinese company — including MiniMax — is legally required to “support, assist, and cooperate with state intelligence work.” The obligation applies continuously and provides no legal pathway for the company to refuse compliance when a government request arrives.

A U.S. congressional investigation announced April 29, 2026 named MiniMax alongside other Chinese AI labs; Anthropic filed February 2026 allegations of industrial-scale distillation against Claude; and an active copyright suit from Disney, Universal, and Warner Bros. Discovery over the Hailuo product was allowed to proceed on May 26, 2026.

The same considerations apply to Kimi K2.7 Code from Moonshot AI (also Beijing-based). The technical merit of these models is real, but enterprise procurement teams in regulated industries, defense-adjacent firms, or any organization handling sensitive IP need to factor these considerations into the decision.

The open-weight nature of both models provides one potential mitigation: if you self-host on your own infrastructure and never send requests to Moonshot’s or MiniMax’s APIs, the data flow to Chinese infrastructure can be cut. Whether that’s sufficient for your compliance requirements is a legal and risk question specific to your organization.

Nemotron Ultra is NVIDIA’s model, an American company, under the Linux Foundation’s OpenMDW-1.1 license. For US enterprises with data sovereignty requirements, that’s a meaningful differentiator that the benchmark tables don’t capture.

---

Final Verdict

Two weeks, three landmark releases. The open-weight AI story in June 2026 isn’t one of gradual improvement — it’s a step change.

MiniMax M3 is the most versatile of the three and the best value proposition in absolute cost terms. Its 1M token context and native multimodality give it a unique profile for visually-driven development work. The caveats are the unverified benchmarks and the data privacy considerations for enterprise use.

Kimi K2.7 Code is the best agentic coding model in the open-weight space right now, period. Beating Claude Opus 4.8 on MCP tool use as an open-weight model is not a minor benchmark win — it’s the gap between “impressive lab demo” and “actually better at what agents do in production.” The 30% token efficiency improvement over K2.6 makes it meaningfully cheaper to run at scale. The caveat: it’s a coding-only specialist, independent benchmarks are still coming, and the Moonshot AI data provenance questions are the same as MiniMax’s.

Nemotron 3 Ultra is the model you deploy when you need a frontier-quality open-weight model you can trust for enterprise, fine-tune with full transparency, run faster than anything else at its intelligence level, and back with the kind of supply-chain credibility that makes legal and procurement teams comfortable. Its 4.8x throughput advantage over Kimi K2.6 is genuinely transformative for production agent workloads, and its harness-agnostic consistency in agent benchmarks is a quality signal that the benchmark tables alone don’t fully convey.

The wider takeaway from this two-week stretch: the frontier for open-weight AI has moved from “almost as good as closed models” to “better than closed models on specific dimensions that matter.” The MCP tool use result from K2.7 is the clearest signal yet. We’re not waiting for open-source to catch up anymore — in certain dimensions, it’s already ahead.

Have you tested any of these models in production? Drop your experience in the comments — especially if you’ve run head-to-head agent sessions on OpenClaw or Hermes. Real-world data points are worth more than any benchmark right now.

For more AI model deep-dives, follow tech.grahammiranda.com.