Can I Run Kimi K3 Locally? 2.8T / 594 GB 1-bit, Not K2.6

Short answer: Kimi K3 is moonshotai/Kimi-K3: 2.8T total / 104B active, native vision, 1M context, MXFP4 MoE. Unsloth UD-IQ1_S is 594 GB and they want ~610 GB total memory. Hub lists ~2.78T parameters (2026-09-10). A used 3090, a 128 GB Mac, and even K2.6’s 350 GB class all miss. Skip or rent. LocalRig has not measured tok/s.

K3 is not K2.6 with a paint job

About 2.8× the total parameters, and Unsloth’s 1-bit file is already larger than K2.6’s 2-bit. Hub: architecture kimi_k3, ~4.2M downloads, 11.3k likes. Unsloth: full precision 1.56 TB disk; Dynamic 1-bit 594 GB (62% smaller in their copy); Dynamic 2-bit 861.3 GB. Lossless Q8 follows MXFP4 experts + BF16 rest.

Unsloth hardware table (accessed 2026-09-10), total memory:

Dynamic 1-bit S	Dynamic 1-bit M	Dynamic 2-bit XXS	Dynamic 2-bit XL	Q8 (lossless)
610 GB	665 GB	726 GB	880 GB	1.6 TB

Quant table on the same page: UD-IQ1_S 594.0 GB, UD-IQ1_M 648.9 GB, UD-IQ2_XXS 711.1 GB, UD-Q2_K_XL 861.3 GB, UD-Q4_K_XL 1,510 GB, UD-Q8_K_XL 1,560 GB.

The sentence about a Mac Studio connected to a 128 GB RAM device is not “K3 fits 128 GB unified.” 594 GB does not fit 128 GB. Read it as extra host RAM, or ignore it as marketing-adjacent. LocalRig treats 610 GB as the 1-bit working floor.

24 GB / 128 GB / 350 GB — all skip

3090 ≈ 4% of 1-bit. 128 GB ≈ 21%. K2.6’s 350 GB 2-bit ≈ 57% of K3 1-bit. You cannot “almost” fit this. Offload to disk is Unsloth’s “it still works, much slower” path — not a LocalRig daily-driver cell.

Browse used RTX 3090 on eBay → for 27B.

Mac Studio on Amazon → for Flash-0731 or GLM-5.3-Flash 3-bit, not K3.

Rent: RunPod, Vast.ai, GPUMart, Vultr GPU Cloud. Unsloth speed cite: ~20 tok/s generation on B200s if it fits, >120 tok/s throughput. Community-cited, B200-only.

Runtime shortfalls

llama.cpp via Unsloth’s fork (vision PR), not a random llama.cpp tag. They document unslothai/llama.cpp checkout kimi-k3-fullsize-vision, mmproj-BF16, UD-IQ1_S. Sampling: temperature 1.0, top_p 0.95 (agentic top_p 1.0). Context up to 1,048,576. Thinking-only. LocalRig will not paste their Unsloth Studio URL that includes loopback addresses.

Ollama: no packet. Do not invent one.

Related: K2.6, K2.7-Code, DeepSeek-V4-Pro.

Unsloth’s warning about other community 1-bit GGUFs (example IQ1_M at 618.9 GB with PPL 54 vs their 594 GB at 2.58) is a quant quality story, not a reason to buy a 3090. Even the “good” 1-bit is 594 GB. DGX Station is the class they name; a used 3090 is not a Station.

Preserve-thinking always on means KV grows with traces you cannot delete. Combined with 1M max context, this is the opposite of a small-card SKU. llama.cpp n_tokens * 160 budget note on their fork is a runtime footgun for operators who already have 610 GB — still irrelevant to 24 GB.

Do not cite Hub likes (11k) as a local-runnability signal. Can I run MiniMax M3 locally? is another 100 GB+ multimodal MoE; K3 is an order of magnitude above that 1-bit floor.

Unsloth is “still investigating if we can push it under 512GiB (dynamic 1-bit is 553.2 GiB)” — even that research floor is four Spark cubes, not one. Do not wait for a 24 GB quant. It will not arrive as a 3090 file.

B200 ~20 tok/s generate / >120 tok/s throughput is the only speed sentence on their page. DGX Station is named. A Mac Studio plus a second 128 GB box is still not 610 GB unless you actually attach that RAM. Read the bill of materials, not the adjective “Mac.”

K3 vs DeepSeek Pro: both are skip/rent. K3 1-bit 594 GB vs Pro Unsloth Q4 ~850 GB. Neither is Flash-0731. If the question is “largest open Moonshot on a home box,” the answer is still none of K2.6 / K2.7 / K3 on 24 GB.

Thinking effort low/high/max changes decode time on a machine that fits. It does not create a 128 GB fit.

Unsloth llama.cpp fork for vision is a build pin. Stock llama.cpp tags may fail. That is an operator cost on a 610 GB box, not a 3090 issue.

NAS for AI model storage if you actually pull 594 GB. Storage is the easy part. Memory is not.

Hub MCP (2026-09-10): moonshotai/Kimi-K3 ~2.78T (2779931.8M), architecture kimi_k3, ~4.2M downloads, ~11.3k likes. Unsloth copy is 2.8T / 104B active. Cite both. Downloads are not a 128 GB fit.

What to run instead

24 GB: Qwen 3.8 27B remains the LocalRig dense daily driver. K3 will not become a 3090 file.

128 GB: DeepSeek-V4-Flash-0731 (~103 GB 3-bit) or GLM-5.3-Flash 3-bit. K3 1-bit at 594 GB is still about 4.6× that class. Unsloth’s “Mac Studio connected to a 128 GB RAM device” line is extra host RAM, not a Studio-only fit.

K2.6 / K2.7-Code class (350–595 GB): those SKUs are still smaller than K3. If you already bought that box for K2.6, K3 1-bit at ~610 GB working memory is a step up, not a free upgrade. Q4 at 1,510 GB is another planet again.

Rent: B200-class or DGX Station-class inventory if you need the vendor thinking model. Unsloth’s only speed sentence is ~20 tok/s generate on B200s if it fits, >120 tok/s throughput. Community-cited, B200-only. Lanes: RunPod, Vast.ai, GPUMart, Vultr GPU Cloud.

Do not wait for a 24 GB K3 quant. Unsloth’s research floor under 512 GiB is still four Spark cubes, not one 3090.

Who this is NOT for

• K2.6 / K2.7-Code shoppers who wanted 350–595 GB, not 594 GB 1-bit on a 2.8T model. K3 is larger.
• 128 GB Mac / Spark buyers. Wrong class.
• People quoting Unsloth B200 20 tok/s as a Mac number.
• Anyone who needs Instant/non-thinking. Unsloth: not supported.
• 24 GB shoppers. Qwen 3.8 27B.

Methodology

• Fit numbers: Unsloth Kimi K3 guide, 2026-09-10 (594 GB UD-IQ1_S, 610 GB working rule, 1,510 GB Q4).
• SKU identity: Hub MCP ~2.78T; Unsloth 2.8T / 104B active, 1M context.
• Speed: Unsloth ~20 tok/s gen on B200 if it fits. Not transplanted.
• LocalRig first-party: none as of 2026-09-10.

Sources

• Unsloth Kimi K3 local guide (accessed 2026-09-10).
• moonshotai/Kimi-K3.
• unsloth/Kimi-K3-GGUF.
• Can I run Kimi K2.6 locally?.
• Can I run DeepSeek-V4-Pro locally?.
• H100 rental price comparison.