Debate over Kimi K3 inference cost: ~20x more expensive than DeepSeek V4.1 Flash

A technical cost debate over Kimi K3's inference costs and bottlenecks played out on Twitter between YouJiacheng and z4y5f3. The core conclusion: K3's decoding phase is compute bound rather than memory bound, and its compute needs far exceed K2/DeepSeek V4.1 Flash, with serving costs estimated at roughly 20x higher.

Confirmed

• YouJiacheng calculated: at 77k context, MLA computation is about 386B; even at maximum MFU, Kimi doesn't have enough GPUs. He said "6x" was only a lower-bound estimate for a second point.
• Decode compute comparison: K3 needs about 1 TFLOPs per token (FP8 equivalent), while K2 V4.1F needs only about 50 GFLOPs; both can be served in a compute-bound state.
• YouJiacheng gave specific KV cache numbers: K3 at 14,400 bytes per token (24 MLA layers), versus 890 bytes for V4.1F.
• He also noted 96-head MLA compute intensity reaches 363 FLOPs/Byte, whether core attention is compute bound is independent of batch size, and FP8 KV matters for avoiding bottlenecks at large batch sizes.
• z4y5f3 corrected his own estimate: he had previously used HF default precision (MLA in BF16, KDA in FP32), computing the KV cache footprint multiple as 40x; after correction, the conclusion remains K3 being about 20x more expensive than V4.1F, and he added that FP8 also runs faster, which doesn't change the overall conclusion.

Unconfirmed

• The upper bound of the original "20–50x" cost gap came from z4y5f3's performance modeling; its KV cache portion has been corrected by the author himself, settling at around 20x, but disagreements over some premises remain under discussion.

Why It Matters

• This debate sharpened the question of "whether K3 is worth deploying" from anecdotal KV cache comparisons down to specific FLOPs, byte counts, and precision assumptions—showing that precision configurations (MLA/KDA default precisions) can significantly change cost-model conclusions.

2026-10-26 ~ 2026-10-26 · 6 related posts

Full story(3 episodes)→

• Episode 1: Kimi K3.1 Tops Terminal-Bench but Cost Charts Draw Criticism(2026-10-24, 4 posts)
• Episode 2: Debate over Kimi K3 serving costs: claims of 25-30x DeepSeek's(2026-10-24, 5 posts)
• Episode 3: Debate over Kimi K3 inference cost: ~20x more expensive than DeepSeek V4.1 Flash(2026-10-26, 6 posts)

• Kimi K3 serving costs 20-50x more than DeepSeek V4.1 Flash, perf modeling argues — z4y5f3 · 2026-10-26
• [source] YouJiacheng breaks down K3 vs V4.1F KV cache: 14400 vs 890 bytes per token — YouJiacheng · 2026-10-26
• [source] Perf modeler revises K3 cost estimate after precision error: still ~20x pricier — z4y5f3 · 2026-10-26
• 96-Head MLA Hits 363 FLOPs/Byte Compute Intensity; FP8 KV Still Can't Escape Compute Bound — YouJiacheng · 2026-10-26
• [source] Kimi K3 Inference Math: 77k-Context MLA Costs ~386B FLOPs-Equivalent, GPUs Not Enough — YouJiacheng · 2026-10-26
• Kimi K3 needs ~1 TFLOPs per decoded token while V4.1F needs only 50 GFLOPs — YouJiacheng · 2026-10-26