MiniMax M3 vs Kimi K3 vs DeepSeek V4 Pro (Jul 2026)
MiniMax M3 vs Kimi K3 vs DeepSeek V4 Pro (Jul 2026)
Three top-tier Chinese open-weight coding models in July 2026: MiniMax M3 (428B parameters, MIT), Kimi K3 (2.8T parameters, Modified MIT — weights July 27), and DeepSeek V4 Pro (1.6T parameters, MIT). Each takes a different position on the cost / quality / hardware trade-off. For most teams, the right answer is a routing setup that uses more than one.
Here is the current pragmatic decision framework, July 18, 2026.
Last verified: July 18, 2026
The Head-to-Head
| Spec | MiniMax M3 | Kimi K3 | DeepSeek V4 Pro |
|---|---|---|---|
| Vendor | MiniMax (Shanghai) | Moonshot AI (Beijing) | DeepSeek (Hangzhou) |
| Released | Jun 1, 2026 (WAIC 2026 highlight) | Jul 16, 2026 (WAIC 2026 debut) | Q2 2026 |
| Total parameters | 428B (MoE) | 2.8T (MoE) | 1.6T (MoE) |
| Active parameters | ~23B | ~50B | ~49B |
| Architecture | MiniMax Sparse Attention (MSA) | Standard MoE (16 of 896 experts) | Standard MoE |
| Context window | 1M tokens (min 512K guaranteed) | 1M tokens | 1M tokens |
| License | MIT | Modified MIT (weights Jul 27) | MIT |
| Multimodal | Native (image + video) | Native (vision) | Text primarily |
| Computer use | Yes | Yes (via Kimi Code) | Yes |
| Input $/MTok (cache-miss) | $0.60 | $3.00 | $0.44 |
| Input $/MTok (cache-hit) | Not tiered | $0.30 | Not tiered |
| Output $/MTok | $2.40 | $15.00 | $0.87 |
| SWE-bench Pro (indep.) | 59.0% | Pending (vendor-only so far) | Behind top open |
| Frontend Code Arena | Strong | Beats Fable 5, GPT-5.6 Sol per vendor claims | Solid |
| Best for | Balanced coding, multimodal | Frontier-adjacent, agentic, long context | Cost-optimized |
When to Use Which
Use DeepSeek V4 Pro when:
- Cost is the primary constraint. At $0.44/$0.87, V4 Pro is roughly 10× cheaper than Sonnet 5 and 7× cheaper than Kimi K3.
- Steady-state high-volume coding agents where marginal cost per task matters most.
- Batch processing — retrieval, extraction, cheap first-pass summarization, code review over large PR volumes.
- Self-hosting on small hardware — V4 Flash variant (284B/13B active) fits on 2× H200 easily.
Use MiniMax M3 when:
- Best independent coding benchmark for open-weight in 2026. SWE-bench Pro 59% and 80.5% on other coding benchmarks.
- Multimodal coding — image + video input in one context, native (V4 Pro is text-primarily).
- Long-context coding — MSA sparse attention keeps compute cost per token low even at 1M context; V4 Pro and K3 pay more compute at long context.
- Free frontier code — Kunalganglani and other reviews highlight M3 as “free 428B model tested” that competes with Claude Code on many tasks.
- Sparse Attention breakthrough cuts compute costs 28× at long contexts without losing quality.
Use Kimi K3 when:
- Highest open-weight quality ceiling. Vendor benchmarks claim K3 beats Fable 5 and GPT-5.6 Sol on frontend code arena; independent verification landing late July 2026.
- Agentic workflows — K3 was designed for long-horizon agentic tasks (navigating large repos, tool use, iterating on runtime feedback).
- You can afford $3/$15 for a top-of-open pick.
- Self-hosting is coming — weights drop July 27, 2026, so plan hardware or GPU cloud rentals for late July.
Cost-Per-Task Comparison (Practical Example)
For a typical Cursor-style coding session with an agent (~15K input tokens, ~5K output tokens per turn, 20 turns):
| Model | Total tokens | Input cost | Output cost | Total per session |
|---|---|---|---|---|
| DeepSeek V4 Pro | 300K in / 100K out | $0.13 | $0.087 | $0.22 |
| MiniMax M3 | 300K in / 100K out | $0.18 | $0.24 | $0.42 |
| Kimi K3 (all cache-miss) | 300K in / 100K out | $0.90 | $1.50 | $2.40 |
| Kimi K3 (50% cache-hit) | 300K in / 100K out | $0.495 | $1.50 | $1.995 |
| Claude Sonnet 5 | 300K in / 100K out | $1.50 | $2.50 | $4.00 |
| Claude Fable 5 | 300K in / 100K out | $3.00 | $5.00 | $8.00 |
| Gemini 3.5 Pro | 300K in / 100K out | $4.50 | $6.00 | $10.50 |
Read: V4 Pro is 18× cheaper than Fable 5, 5× cheaper than Sonnet 5. MiniMax M3 is 20× cheaper than Fable 5. Kimi K3 sits between Sonnet 5 and Fable 5 on cost while claiming higher-than-Fable 5 quality.
Head-to-Head on Real Workflows
Cursor-style coding agent (many turns, moderate complexity):
- Best: MiniMax M3 (cost + quality balance, MSA long-context efficiency).
- Second: DeepSeek V4 Pro (better cost, small quality gap).
- Frontier: Kimi K3 for hard tasks.
Full-repo refactor (1M+ token context needed):
- Best: MiniMax M3 or Kimi K3 (both handle 1M cleanly).
- For maximum context (>1M): Gemini 3.5 Pro (2M) — not open-weight.
Multimodal code review (screenshots, diagrams, video walkthroughs):
- Best: MiniMax M3 (native multimodal, cheapest).
- Second: Kimi K3 (native vision).
- Avoid: DeepSeek V4 Pro (text-primary).
Batch code analysis (lots of files, no interactive turns):
- Best: DeepSeek V4 Pro or V4 Flash (cheapest).
- Second: MiniMax M3 (still cheap).
- Skip Kimi K3: cost doesn’t justify the batch use case.
Self-hosted enterprise deployment (data residency, custom fine-tuning):
- Best: MiniMax M3 (MSA gives lowest compute-per-token, easiest to serve).
- Second: DeepSeek V4 Pro or V4 Flash (smallest active parameter count).
- Third: Kimi K3 (biggest, most GPU-intensive; wait for July 27 weights).
The Routing Playbook
Most sophisticated teams in July 2026 route across multiple open-weight models:
Tier 1 (default for 70% of tasks): DeepSeek V4 Pro or MiniMax M3 — whichever is cheaper for your token profile.
Tier 2 (25% of tasks that need higher quality): MiniMax M3 (if not already tier 1) or Claude Sonnet 5.
Tier 3 (5% of hard tasks): Kimi K3, Claude Fable 5, or Gemini 3.5 Pro depending on task type.
Router tooling: OpenRouter, Kilo.ai, Aider, Continue, and Cursor all support multi-model routing via config. Set up cost-per-token-aware routing so cheap models handle the load and expensive models only get triggered on task complexity signals.
Sub-Questions People Are Asking
Is MiniMax M3 really the best open-weight coder? As of July 18, on independent SWE-bench Pro, yes at 59.0%. Kimi K2.6 was 58.6% — close but behind. Kimi K3 could leapfrog once independent benchmarks land in late July.
Are these models safe for enterprise use? For technical capability: yes. For legal/compliance: MIT-family licenses are permissive. For political-risk / provenance: Chinese-origin models may not clear US federal or some EU regulated financial procurement — this depends on your specific compliance posture. Self-hosting on your own infrastructure addresses data residency but not model-provenance policies.
Should I self-host or use the API? Use API until you’re doing more than ~100M tokens/day steady traffic. Above that, self-host on GPU cloud or your own hardware. See our Kimi K3 self-hosting guide for detailed break-even math.
Can I fine-tune these? All three ship weights. All three can be full or LoRA fine-tuned with torchtune, unsloth, or MegaBlocks. MoE fine-tuning at 428B (M3), 1.6T (V4 Pro), or 2.8T (K3) requires serious GPU budget — expect $30-200K for a serious fine-tune depending on model and dataset size.
What about GLM-5.2, Qwen 3.6? GLM-5.2 (from Zhipu AI, ~380B) and Qwen 3.6 Plus (from Alibaba, ~750B) are also strong Chinese open-weight options but sit slightly behind M3/K3/V4 Pro on current benchmarks. Include them in your routing consideration but they aren’t the top three.
What’s Actually Different About Each
MiniMax M3 differentiator: MSA (MiniMax Sparse Attention) architecture makes long-context compute cheap. This translates directly into low API pricing and easier self-hosting. Best price/performance in mid-tier.
Kimi K3 differentiator: Largest total parameter count of any open-weight model, designed specifically for long-horizon agentic workflows. Best raw capability ceiling — if vendor benchmarks hold up in independent testing.
DeepSeek V4 Pro differentiator: Aggressive cost optimization. V4 Flash variant (284B/13B active) is the cheapest capable coding model on the market at $0.14/$0.28. Best pure cost.
Bottom Line
In July 2026, the Chinese open-weight coding ecosystem is genuinely competitive with frontier proprietary models on the majority of coding tasks. MiniMax M3 is the current benchmark leader, DeepSeek V4 Pro is the cost leader, and Kimi K3 is the potential frontier leader pending independent benchmarks.
Practical July 2026 recommendation:
- Cost-optimized default: DeepSeek V4 Pro.
- Balanced quality/cost default: MiniMax M3.
- Best open-weight ceiling: Kimi K3 (after Jul 27 weight drop and independent benchmark verification).
- Fallback for hardest tasks: Claude Fable 5, GPT-5.6 Sol, Gemini 3.5 Pro (proprietary).
Set up routing across at least two of these. Chinese open-weight is now a serious cost lever without meaningful quality sacrifice on routine tasks — the story of AI coding in the second half of 2026.
Sources
- MiniMax M3 official launch specs: minimax.io/models/text/m3
- MiniMax M3 architecture + benchmarks: morphllm.com/minimax-m3
- Kimi K3 model card and license: kimi.com/blog/kimi-k3
- Kilo.ai open-source coding models 2026: kilo.ai/open-source-models
- Independent coding leaderboard: kilo.ai/leaderboard