DeepSeek V4.1 Flash vs Kimi K3: The Open-Weight Duel

Kimi K3 and DeepSeek V4.1 Flash are both open-weight frontier models with 1M-token contexts, native multimodality, and agent-oriented APIs — but they are built to different scales. K3 is a 2.8-trillion-parameter MoE with 104B active parameters; V4.1 Flash is a 552B backbone with 8B prefill and 16B decode activation^[4][5].

K3's differentiator is always-on deep reasoning plus native image and video understanding, wrapped in a bespoke license. V4.1 Flash's is extreme inference efficiency: a Causal Encoder-Decoder core, an Engram memory, and an 890-byte-per-token KV cache that lets DeepSeek charge a fraction of K3's rates^[2][4][5].

Both are credible frontier options, which is exactly why the comparison is interesting. It is no longer a case of a small open model against a giant one — it is two very large models with sharply different cost and licensing trade-offs^[4][5].

On GPQA Diamond, a graduate-level science QA set, Kimi K3 scores 92.9 against V4.1 Flash's 90.9 — a two-point edge to K3. Both trail GPT-5.6 Sol (94.1) and Opus 5 (93.4), so this is a case where the two open models are close and the closed frontier is still slightly ahead^[4].

GPQA is a knowledge-and-reasoning proxy rather than an agentic test. A two-point gap rarely changes a deployment decision on its own; it signals that K3 has a small edge on hard scientific questions answered from knowledge^[4].

The more useful reading is that V4.1 Flash is within two points of K3 while costing roughly 20-25x less on uncached tokens and 100x less on cache hits. For high-volume question answering, that trade favors DeepSeek unless the last two points are decisive^[3][4][5].

Humanity's Last Exam widens the gap. Kimi K3 scores 43.5 versus V4.1 Flash's 36.8, a 6.7-point lead on the hardest closed-book reasoning set. With tools the two converge — and actually flip: V4.1 Flash scores 63.9 and K3 scores 59.8^[4].

That inversion is the most important detail in the comparison. When a model can retrieve evidence and run tools, DeepSeek's deficit on closed-book HLE largely disappears and becomes a small lead. When it must answer from parametric knowledge alone, K3 is clearly stronger^[4].

For agent builders this argues for evaluating the tool-assisted row, not the closed-book row. The gap that matters is whichever mode your production system actually uses^[4].

On Terminal-Bench 2.1, an established agentic coding benchmark, V4.1 Flash leads 90.6 to 88.3. It also wins the newer Terminal-Bench 3.0 (30.0 vs 17.7) and Terminal-Bench 4.0 (31.2 vs 12.6), so the terminal-agent comparison breaks in DeepSeek's favor at every version DeepSeek reports^[4].

The TB 4.0 margin is the striking one: 31.2 against 12.6. That is a 18.6-point lead on the hardest terminal benchmark, which suggests V4.1 Flash handles long-horizon shell and systems tasks materially better than K3 in DeepSeek's runs^[4].

The usual caveat applies. These are DeepSeek-run evaluations using DeepSeek Harness at maximum effort, and K3's own vendor materials may tell a different story on different harnesses. Treat TB as a strong hint, not a verdict^[4][5].

DeepSWE v1.1 goes to V4.1 Flash 74.2 to 67.5 — a 6.7-point gap on software-engineering tasks. CyberGym goes 88.1 to 80.0 on security agent tasks, and AutomationBench goes 54.8 to 46.7. These are the agentic-coding rows most relevant to coding agents^[4].

The consistency is notable: across coding, security, and automation agent benchmarks, V4.1 Flash leads K3 by 6-8 points. Combined with the terminal results, DeepSeek's own table paints a clear picture of V4.1 Flash as the stronger agentic coder^[4].

That said, benchmark position is not cost per completed task. K3's 104B active parameters may reduce error rates on tasks it attempts, and its always-on thinking can help on partly specified problems. The only reliable test is your own workload with your own harness^[4][5].

• DeepSWE v1.1: 74.2 vs 67.5 (V4.1 Flash +6.7)
• CyberGym: 88.1 vs 80.0 (V4.1 Flash +8.1)
• AutomationBench: 54.8 vs 46.7 (V4.1 Flash +8.1)
• NL2Repo-Bench: 64.0 vs 58.0 (V4.1 Flash +6.0)
• ProgramBench: 20.3 vs 17.5 (V4.1 Flash +2.8)^[4]

Here the models diverge sharply. V4.1 Flash is $0.003 per million cached input tokens, $0.15 uncached, and $0.60 output off-peak. Kimi K3, per Moonshot's July 25, 2026 price page, is $0.30 cached input, $3.00 uncached input, and $15.00 output — a cache-hit gap of about 100x and roughly 20-25x on uncached input and output^[3][5].

A concrete request shows the scale: one million uncached input tokens plus 200,000 output tokens costs about $0.27 on V4.1 Flash at off-peak rates versus roughly $6.00 on K3. If the whole prefix qualifies for the cache-hit rate, the same request is fractions of a cent on DeepSeek and about $3.30 on K3^[3][5].

Peak rates double V4.1 Flash's prices but do not close a 20-100x gap. For token-heavy agent pipelines, pricing is the single biggest practical difference between these two models^[3].

Both publish weights, but the terms differ. V4.1 Flash is MIT-licensed on Hugging Face. Kimi K3 uses a bespoke license with attribution and commercial-use conditions that large-scale consumer services and model-as-a-service businesses must examine. If permissive reuse matters, DeepSeek is the simpler choice^[4][5].

Self-hosting costs are large for both. V4.1 Flash needs roughly 511 GB of disk and about 614 GB of VRAM, meaning a GB200 NVL4 tray or eight H200s. K3's 2.8T parameters make it heavier still. Open weights buy control and auditability, not cheap inference^[4][5].

For startups and research groups, the MIT license removes legal friction and permits fine-tuning and redistribution. For companies planning to resell model access at scale, K3's license terms require legal review before launch^[4][5].

When V4.1 Flash launched, its combination of frontier scores and low prices hit Chinese AI peers: MiniMax and Z.ai shares in Hong Kong fell more than 8%, and Alibaba dropped more than 2%. The read is that a cheap open-weight model at this capability level resets what competitors can charge^[4][6].

Choose V4.1 Flash for cost-sensitive coding and security agents, cache-heavy pipelines, and any deployment that needs MIT-licensed self-hosting. Choose Kimi K3 for workflows that need native video understanding, always-on thinking, or closed-book reasoning where its HLE lead holds^[4][5].

As always, run your own evaluation. Both sets of scores are vendor-reported, the agentic rows depend on the harness, and a 6-point benchmark lead can vanish or invert on a task distribution the vendors did not test^[4][6].

• High-volume coding agents → V4.1 Flash; 6-8 point lead at a fraction of the price.
• Native video input → Kimi K3; V4.1 Flash handles images, not video.
• Strict permissive licensing → V4.1 Flash (MIT).
• Always-on deep reasoning on under-specified tasks → test K3 against your prompts.
• Cache-heavy repetitive pipelines → V4.1 Flash; the cache-hit gap is ~100x^[3][4]