Evidence reviewed July 31, 2026. This is a source-based comparison with a separately labeled ChatLLM observation, not a hands-on official-API review or a locally reproduced hardware benchmark.
The short answer
Kimi K3 is the source-based capability pick if you care most about the July 31 Artificial Analysis Intelligence Index v4.1 result and documented image input; DeepSeek V4 Flash 0731 is the value and local-experimentation pick if first-party API cost or home-lab feasibility matters more.
That verdict needs four guardrails. In Artificial Analysis’s proprietary, dynamic Intelligence Index v4.1 snapshot captured 2026-07-31 PDT, its max-reasoning entries scored Kimi 57 and first of 101 models, versus DeepSeek at 50 and third. DeepSeek’s first-party list prices were dramatically lower under every explicit workload illustration in this comparison. Community projects document quantized DeepSeek-family execution on high-memory Macs and a version-pinned 0731 deployment on two DGX Sparks, while Kimi’s official artifact and deployment guidance remain cluster-class. Kingy did not obtain a successful completion from either official API or reproduce either model on Mac Studio or DGX Spark hardware.
There is no evidence-supported universal winner. The correct choice depends on whether you value a named independent composite, API economics, image input, hosted-provider breadth or a community-supported route to local inference.
Choose Kimi K3 if
- You prioritize Kimi’s 57-to-50 lead on Artificial Analysis’s proprietary, dynamic Intelligence Index v4.1 max-reasoning snapshot captured 2026-07-31 PDT.
- You need a checkpoint whose official model card lists text and image input.
- You prefer the broader hosted-provider coverage visible in Artificial Analysis’s dated tracker, accepting that tracker coverage is not a regional availability guarantee.
Choose DeepSeek V4 Flash 0731 if
- First-party list price is a primary constraint.
- You want the only credible documented home-lab path in this comparison, while accepting that it is community-supported, quantized and not reproduced by Kingy.
- You want a much smaller official repository footprint, without treating smaller storage as proof of equal quality or faster inference.
Choose neither yet if
- Your purchase or deployment depends on independently measured official latency, reliability, tool-use success, token accounting or long-context behavior; Kingy did not obtain those measurements.
- Your decision depends on a particular coding agent or tool workflow and you cannot run a workload-specific evaluation, because the vendors’ similarly named benchmark rows were produced with non-equivalent harnesses and settings.
First, untangle the model names
Kimi K3 is an official Moonshot AI model, and its first-party API identifier is kimi-k3. DeepSeek’s current documentation uses the alias deepseek-v4-flash and maps it to DeepSeek-V4-Flash-0731. That is a documentation-backed mapping, not a live-backend identity Kingy independently authenticated.
The version distinction matters because the 0731 release supersedes DeepSeek’s April preview through additional post-training while retaining the preview architecture and size according to DeepSeek’s changelog. Preview-era architecture material can therefore explain the family, but preview scores must not be substituted for 0731 results.
The two parameter stories are also easy to misread. Moonshot describes Kimi K3 as a 2.8-trillion-parameter mixture-of-experts model with approximately 104 billion active parameters. “Active” refers to parameters selected during inference, not the size of the checkpoint that must be stored or loaded.
DeepSeek’s preview architecture is described as 284 billion total and 13 billion active parameters, while the current 0731 Hugging Face page displays 304 billion parameters and separately documents attached speculative decoding. The reviewed official sources do not reconcile those differently scoped figures, so this comparison does not claim that the speculative component causes the difference.
| Specification | Kimi K3 | DeepSeek V4 Flash 0731 |
|---|---|---|
| Official API name | kimi-k3 |
Official documentation maps deepseek-v4-flash to DeepSeek-V4-Flash-0731; live mapping not independently authenticated |
| Architecture figures | 2.8T total / about 104B active | Preview architecture: 284B total / 13B active; current 0731 page displays 304B |
| Documented input | Official checkpoint card: text and image | Text only |
| Context | 1,048,576 tokens | 1M tokens |
| Maximum output | max_completion_tokens defaults to 131,072 and has a documented 1,048,576 ceiling; maximum input and output are not proven to fit simultaneously |
384K documented maximum output |
| Reasoning controls | Always reasons; low, high and max, default max |
Thinking and non-thinking modes; low, high and max documented |
Kimi’s checkpoint-card modality is the cleanest comparison point: text and image for Kimi versus text only for DeepSeek 0731. Moonshot’s product and API materials also discuss uploaded video, but those materials describe modality differently from the downloadable checkpoint card; this article therefore does not turn product-level video handling into a native checkpoint claim.
What Kingy observed—and what we did not verify
This is a source-based comparison with supplemental third-party platform observations, not a controlled benchmark of the official checkpoints.
Kingy’s official-endpoint attempts ended in authentication or billing errors and produced no successful model completion. Those failures describe the attempted access paths, not either model’s quality. Neither provider returned a served-model identifier, system fingerprint, usage record, completion, reasoning output, tool call, context result or billable-token cost.
No Mac Studio or DGX Spark was available to this project, and no local Kimi or DeepSeek checkpoint was loaded. Every local-performance number below is consequently labeled as an official hardware specification, a community report or a Kingy inference from raw capacity—not as our benchmark.
Supplemental ChatLLM observation
In the signed-in ChatLLM interface, each visible model option was tested with 11 one-shot cases in fresh conversations, and both were semantically correct on 11 of 11. Kimi was visibly format-compliant on 11 of 11, while DeepSeek was compliant on 10 of 11 because one correct Python answer appeared with extra material; that visible difference may come from the model, ChatLLM’s rendering, or both.
ChatLLM displayed 531 credits for Kimi and 22 for DeepSeek, or 553 total, but the result views did not document a conversion to tokens or dollars. Those opaque credits are a ChatLLM subscription observation, not evidence about official API price or token efficiency. The UI labels also did not expose exact served checkpoints, so these results are not proof that ChatLLM routed the official Kimi K3 and DeepSeek V4 Flash 0731 weights.
Architecture, context and reasoning controls
Kimi K3 uses Kimi Delta Attention, Attention Residuals and Stable LatentMoE, with 896 experts and 16 selected per token according to Moonshot’s official materials. Those are vendor-documented architecture facts, not independent proof of speed or quality.
Both models advertise roughly one-million-token context ceilings. Kimi’s documented context is 1,048,576 tokens, while DeepSeek’s is 1M. Kimi’s max_completion_tokens defaults to 131,072 and has a documented ceiling of 1,048,576, but that does not establish that maximum input and maximum output can be combined in one request. DeepSeek documents a 384K maximum output. Kingy did not exercise the maximum usable context or output length on either service.
Reasoning controls differ in ways that complicate a supposedly identical API benchmark. Kimi always reasons and supports low, high and max, defaulting to max. DeepSeek supports thinking and non-thinking modes and documents low, high and max reasoning effort for the current 0731 materials. Kimi documents fixed temperature 1.0 and top-p 0.95 behavior, while DeepSeek thinking mode documents sampling controls such as temperature and top-p as unsupported or no-op. That interface asymmetry is why the aborted official harness could not honestly claim exact sampling parity, and it says nothing about the settings ChatLLM used behind its interface.
Benchmarks: one common composite, several non-common vendor tables
The strongest common comparison in the closed evidence base is Artificial Analysis Intelligence Index v4.1. In its 2026-07-31 PDT snapshot, the site’s max-reasoning entries scored Kimi K3 at 57 and DeepSeek V4 Flash 0731 at 50. Kimi ranked first of 101 and DeepSeek ranked third of 101 in that snapshot. Artificial Analysis uses a proprietary, dynamic methodology, so the result should be read as a named and dated composite—not a timeless universal ranking.
Artificial Analysis also reports that its index evaluation generated about 130 million output tokens and cost $2,437.41 for Kimi, versus about 210 million output tokens and $72.02 for DeepSeek. Those figures are observations about Artificial Analysis’s evaluation run, verbosity and model pricing; they are not additional quality scores or a prediction of an individual user’s bill.
The vendor tables are a different evidence class. All figures below are vendor-reported and were captured 2026-07-31 PDT. Kimi’s official table uses max reasoning, temperature 1.0 and benchmark-dependent top-p and harness settings. DeepSeek’s public 0731 code-agent results use DeepSeek Harness minimal, max effort, temperature 1.0 and top-p 0.95, while the harness was announced as forthcoming. A higher number in either column can be directional, but it is not an apples-to-apples win.
| Benchmark | Kimi K3 vendor result | DeepSeek V4 Flash 0731 vendor result | Comparability |
|---|---|---|---|
| Terminal-Bench 2.1 | 88.3 | 82.7 | Directional only; different harnesses and top-p. |
| DeepSWE | 67.5 in Kimi’s table; 67.3 in its footnote | 54.4 | Do not rank from the raw values because Kimi’s source conflicts internally and the harnesses differ. |
| Toolathlon Verified | 76.5 | 70.3 | Directional only; tool and harness setups differ. |
| Agents’ Last Exam | 28.3 | 25.2 | Same-name benchmark, not a common run. |
| AutomationBench | 30.8 | 25.1 (Public) | Public subset, version and harness alignment are not proven. |
All Kimi figures in that table are vendor-reported for Kimi K3 under Moonshot’s official table, with Kimi Code, max reasoning, temperature 1.0 and top-p 1.0 for the agentic rows. The DeepSeek figures are vendor-reported for DeepSeek V4 Flash 0731 under DeepSeek Harness minimal, max effort, temperature 1.0 and top-p 0.95; DeepSeek announced that harness as forthcoming. Terminal-Bench therefore compares Kimi Code at top-p 1.0 with DeepSeek Harness minimal at top-p 0.95. In DeepSWE, Kimi’s table reports 67.5 under Kimi Code while its footnote points to 67.3 under mini-SWE-agent; DeepSeek reports 54.4 under its 0731 minimal-harness settings. Toolathlon uses Kimi Code versus DeepSeek Harness minimal, both at max reasoning and temperature 1.0 but with top-p 1.0 versus 0.95. Agents’ Last Exam uses Kimi Code versus DeepSeek Harness minimal under those same vendor settings, so it is a same-name rather than common run. AutomationBench reports Kimi’s public subset under Kimi Code and DeepSeek’s Public row under DeepSeek Harness minimal, but dataset/version alignment is not proven.
Kimi K3’s other vendor-only rows, captured 2026-07-31 PDT from Moonshot’s official table, include GPQA Diamond 93.5, ProgramBench 77.8, FrontierSWE 81.2 and BrowseComp 91.2 under max reasoning, temperature 1.0 and benchmark-dependent harness/top-p settings. Moonshot also reports BrowseComp 90.4 when using the full 1M context without compaction, so the context-management distinction is part of that vendor result.
DeepSeek V4 Flash 0731’s other vendor-only rows, captured 2026-07-31 PDT under DeepSeek Harness minimal, max effort, temperature 1.0 and top-p 0.95, include NL2Repo 54.2 and Cybergym 76.7, plus internal DSBench-FullStack 68.7 and DSBench-Hard 59.6. The DSBench rows are internal and should not be treated as independently auditable results from the present evidence.
The benchmark conclusion is narrow: Kimi leads Artificial Analysis’s proprietary, dynamic Intelligence Index v4.1 max-reasoning snapshot captured 2026-07-31 PDT, while the vendor tables indicate areas worth testing but cannot establish a general winner.
API access, pricing and provider availability
Moonshot documents an OpenAI-compatible Kimi endpoint through its Kimi K3 quickstart, using the model alias kimi-k3. As captured 2026-07-31 PDT, the same documentation says K3 is unlocked after a successful minimum $1 top-up and that cumulative top-up affects rate limits.
DeepSeek documents both OpenAI- and Anthropic-format access and uses the deepseek-v4-flash alias. Its current documentation maps that alias to 0731, but Kingy did not independently authenticate the live backend. DeepSeek describes V4 Flash 0731 as a public-beta API, which is not a guarantee that every region, account or billing state has access.
First-party list prices captured 2026-07-31 PDT
| Price per 1M tokens | Kimi K3 | DeepSeek V4 Flash 0731 |
|---|---|---|
| Cache-hit input | $0.30 | $0.0028 |
| Cache-miss input | $3.00 | $0.14 |
| Output | $15.00 | $0.28 |
The Kimi prices above are current first-party list rates captured on 2026-07-31 PDT and exclude taxes and provider markups. The DeepSeek prices are the corresponding first-party list rates captured on the same date. DeepSeek separately says a future peak/off-peak policy will charge twice the base rates during stated Beijing-time peak windows, but the reviewed page gives no effective date; the surcharge is therefore not applied to the base calculations below.
Transparent workload illustrations
These are arithmetic illustrations using first-party list prices, not universal bill estimates. They exclude taxes, provider markups, hidden or tool-call tokens and DeepSeek’s announced future surcharge. The Artificial Analysis 7:2:1 blend orders cache-hit input, cache-miss input and output tokens in that ratio.
| Workload assumption | Kimi K3 | DeepSeek V4 Flash 0731 | Relative listed cost |
|---|---|---|---|
| 1M cache-miss input + 100K output | $4.50 | $0.168 | Kimi about 26.8x higher. |
| 1M cache-hit input + 100K output | $1.80 | $0.0308 | Kimi about 58.4x higher. |
| Artificial Analysis 7:2:1 blend | $2.31 per 1M blended tokens | $0.05796 per 1M blended tokens | Kimi about 39.9x higher for this workload illustration. |
This is the clearest DeepSeek advantage in the comparison: under the explicit scenarios above, its official list pricing is far lower. It is still better editorial practice to show the workload than to say DeepSeek is “cheaper in every circumstance,” because taxes, caching, hidden tokens, provider routing and future pricing policies can change the bill.
Artificial Analysis tracked 10 Kimi providers and one DeepSeek 0731 provider in its 2026-07-31 PDT snapshot. That is tracker coverage, not proof of availability in every region or account, nor proof that every mirror routes an identical checkpoint.
Under Artificial Analysis’s 2026-07-31 PDT 10K-input provider snapshot, Wafer (FAST) led tracked Kimi providers at 172 output tokens per second, while first-party Kimi showed 35. The tracker had no DeepSeek 0731 provider speed or latency data, so those Kimi figures cannot support a Kimi-versus-DeepSeek speed comparison.
Can either model run on a Mac Studio or DGX Spark?
The answer changes sharply depending on whether “run” means the pinned official repository artifact, an official deployment recipe or an aggressive community quantization. Repository bytes are useful capacity evidence, but they are not loaded-memory requirements, KV-cache capacity or proof of usable speed.
At the pinned Kimi Hugging Face revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569, repository storage was 1,561,018,243,668 bytes. At the pinned DeepSeek 0731 revision 7872f01b1d1fe23eabc4c98b48bffcef5a386062, repository storage was 166,886,535,336 bytes. Those values describe repository-level storage at capture, not the complete runtime working set.
Moonshot recommends Kimi deployment on supernodes with 64 or more accelerators; that is official guidance, not a minimum requirement Kingy validated. DeepSeek’s official 0731 model-card example uses one node with four GB300 accelerators, vLLM and DSpark speculative decoding, which is an official data-center example rather than a home-desktop recipe.
Apple documents current Mac Studio memory up to 128GB on M4 Max and up to 512GB on M3 Ultra, with 819GB/s memory bandwidth on M3 Ultra. Regional configurations can change. NVIDIA documents 128GB unified memory and 273GB/s bandwidth per DGX Spark, plus a generic parameter envelope up to 200 billion on one Spark or 405 billion across two. That generic parameter envelope does not prove that a particular checkpoint fits with useful context or speed.
As a raw-capacity comparison, Kimi’s pinned 1.56TB repository footprint does not fit the listed 128GB or 512GB Mac Studio tiers or the 128GB/256GB capacity of one or two DGX Sparks. That does not prove that no future extreme quantization, streaming method or different artifact could ever work.
Hardware feasibility matrix
| Hardware | Kimi K3 | DeepSeek V4 Flash 0731 | Evidence verdict |
|---|---|---|---|
| Mac Studio M4 Max, 128GB | Pinned official footprint exceeds raw capacity. | Community q2 class targets 96/128GB machines. | DeepSeek community path only; no Kingy local result. |
| Mac Studio M3 Ultra, 256GB | Pinned official footprint exceeds raw capacity. | Community q4 guidance starts at 256GB. | DeepSeek local-experiment candidate; not reproduced. |
| Mac Studio M3 Ultra, 512GB | Pinned official footprint exceeds raw capacity. | Community q4 Flash-family execution is documented. | Feasibility evidence, not a Kingy 0731 benchmark. |
| One DGX Spark, 128GB | Pinned official footprint exceeds raw capacity. | Community q2 Flash-family execution is documented. | Aggressive community quantization path only. |
| Two DGX Sparks, 256GB aggregate | Pinned official footprint still exceeds raw capacity. | Version-pinned community TP=2 0731 recipe. | Experimental, self-reported and unreproduced. |
DwarfStar’s 512GB Mac Studio row reports 26.62 tok/s for q4 generation with a 12,018-token prompt, --ctx 32768, no-thinking, greedy decoding and 256 generated tokens. It is a self-reported single Flash-family run, may be stale and does not pin the checkpoint to 0731. Its one-DGX-Spark row reports 13.75 tok/s for q2 generation with a 7,047-token prompt, --ctx 32768, no-thinking, greedy decoding and 256 generated tokens. That result is also self-reported, potentially stale, not checkpoint-pinned to 0731 and not a Kingy benchmark.
The two-Spark recipe at commit 914c35bd7d5607560048e4467c3fdd42e892e297 documents a TP=2 0731 deployment with NVFP4 KV cache, a configured 1M maximum length and a hash-pinned 19-case artifact. The recipe is experimental, self-reported and unreproduced, and its configured ceiling does not establish reliable full-length performance under every load.
DeepSeek is therefore the only model here with a credible documented home-lab path in the closed evidence base. “Credible documented path” is intentionally narrower than “easy,” “practical for everyone” or “tested by Kingy.”
Licenses and commercial-use traps
Kimi K3 is open-weight under the custom Kimi K3 License. It should not be described as unrestricted MIT- or Apache-style open source.
The license says certain Model-as-a-Service operators above $20 million in aggregate revenue over a consecutive 12-month period must enter a separate agreement with Moonshot before commercial use. That condition applies to the defined MaaS scope and threshold; it is not a blanket ban on commercial use.
The license also requires certain commercial products above 100 million monthly active users or $20 million in monthly revenue to display “Kimi K3” prominently, with stated exceptions for internal use and official or certified access. Organizations near those thresholds should obtain their own legal review rather than rely on this summary.
DeepSeek’s official 0731 repository and model weights are MIT-licensed. That statement is deliberately scoped to the exact official distribution. Community GGUFs, runtime code, containers, DSpark files and hosted providers can carry separate terms, so each artifact needs its own license check before redistribution or commercial deployment.
This section is a source summary, not legal advice.
The practical dividing line
Kingy’s source-based recommendation is Kimi when the dated Artificial Analysis lead, documented image input and broader tracked hosted availability matter more than API price or local deployment. The image-input advantage does not establish that Kimi is better at every visual task. DeepSeek is the better fit when captured first-party list price or a community-supported home-lab path is the primary constraint, while its smaller repository footprint does not prove equal quality or faster inference.
Do not choose either model solely from its vendor table if the decision depends on a specific coding agent, tool workflow, latency target or long-context workload. A buyer should validate that workload with declared prompts, tools, reasoning controls and scoring because the frozen Path B evidence contains no successful common official-endpoint run.
What remains unknown
The live first-party DeepSeek backend was not independently fingerprinted as 0731. Exact official endpoint latency, throughput, reliability, tool-use success and token accounting remain unmeasured by Kingy for both models. Maximum usable context and output lengths were not exercised.
The official sources reviewed do not reconcile DeepSeek’s 284B/13B preview-architecture figures with the 304B figure displayed on the current repository page. Kimi’s checkpoint card and its product/API materials describe modality differently, particularly around video.
Provider counts, performance, aliases and prices can change after the 2026-07-31 PDT capture. No project-owned Mac Studio or DGX Spark result exists.
Those are not footnote-sized limitations. They define what this comparison can tell a technical buyer and where the evidence stops.
FAQ
Is Kimi K3 better than DeepSeek V4 Flash 0731?
Kimi led Artificial Analysis’s proprietary, dynamic Intelligence Index v4.1 max-reasoning snapshot by 57 to 50 on 2026-07-31 PDT, but that named composite does not establish universal superiority. Kingy’s conditional recommendation is Kimi for the dated capability signal and image input, DeepSeek for official API value and local experimentation, and neither without workload-specific validation when latency or agent behavior is decisive.
Which model is cheaper through its official API?
As captured 2026-07-31 PDT, DeepSeek’s first-party list rates were $0.0028/M cache-hit input, $0.14/M cache-miss input and $0.28/M output, versus Kimi’s $0.30/M, $3.00/M and $15.00/M respectively. The exact bill depends on the workload, and DeepSeek separately announced a future twice-base-rate peak policy without an effective date on the reviewed page.
Can Kimi K3 run on a Mac Studio?
Kimi’s pinned official repository footprint was about 1.56TB, Moonshot recommends supernodes with 64 or more accelerators, and current Mac Studio tiers in the evidence top out at 512GB. As a raw-capacity comparison, the pinned artifact does not fit those machines, but this article does not claim that no future extreme quantization or streaming approach could ever work.
Can DeepSeek V4 Flash run on a Mac Studio?
Community DwarfStar/ds4 guidance documents q2 for 96/128GB machines and q4 for machines with at least 256GB. Its cited 512GB Mac speed result is a self-reported Flash-family run whose checkpoint is not pinned to 0731, so it is feasibility evidence rather than a Kingy benchmark of the named release. DeepSeek nevertheless has the only credible documented home-lab path in this comparison.
Can DeepSeek V4 Flash run on one DGX Spark?
One DGX Spark has 128GB unified memory, while the pinned official DeepSeek repository storage was about 166.9GB. DwarfStar reports a q2 community path at 13.75 tok/s with a 7,047-token prompt, --ctx 32768, no-thinking, greedy decoding and 256 generated tokens. The result is self-reported, potentially stale, not checkpoint-pinned to 0731 and not a Kingy benchmark.
What about two DGX Sparks?
NVIDIA documents a generic envelope up to 405 billion parameters across two Sparks. A version-pinned community recipe documents a two-Spark 0731 deployment with a configured 1M maximum length, but it is experimental, self-reported and unreproduced; the configured ceiling does not prove reliable full-context operation under all loads.
Do both models support one-million-token context?
Kimi documents 1,048,576 tokens of context, while DeepSeek documents 1M. Those are documented ceilings, not lengths Kingy successfully exercised, and Kimi’s maximum input and maximum completion ceilings should not be assumed to fit simultaneously.
Are both models open source and commercially usable?
Kimi K3 is open-weight under a custom license with threshold-based MaaS and branding conditions. DeepSeek’s official 0731 repository and weights are MIT-licensed, but third-party quantizations, runtimes, containers and hosted services can have separate terms.
What did Kingy actually test?
Kingy’s official-endpoint attempts returned access or billing errors and no successful completion, so they produced no official model-quality, latency, throughput or token-accounting result. The supplemental ChatLLM check consisted of 11 fresh one-shot cases per visible model option, with both semantically correct on 11 of 11; the platform did not expose exact served checkpoints or a token/dollar conversion for its credits. No Mac Studio or DGX Spark local run was performed.
Final verdict
Kimi has the stronger source-based capability case: it leads Artificial Analysis’s proprietary, dynamic Intelligence Index v4.1 max-reasoning snapshot captured 2026-07-31 PDT, and its official checkpoint card documents image input. DeepSeek has the stronger economic and local-experimentation case: its first-party list prices are far lower in the explicit scenarios, and community projects document the only credible home-lab route in this comparison.
That split is more useful than declaring an overall winner. The vendor benchmark rows are not common runs, the official APIs were not successfully benchmarked by Kingy, and no local hardware result was reproduced. Readers who need verified latency, reliability, tool use or full-context behavior should treat the comparison as incomplete until they validate their own workload.
Sources
Official Kimi sources
- Kimi K3 technical blog — architecture, pricing and deployment guidance.
- MoonshotAI/Kimi-K3 GitHub repository and official Hugging Face model card — architecture, context, modality, benchmarks and checkpoint metadata.
- Kimi K3 License — commercial and branding conditions.
- Kimi K3 API quickstart and reasoning-effort guide — access, context and reasoning controls.
Official DeepSeek sources
- DeepSeek API changelog — 0731 release identity and vendor benchmark settings.
- DeepSeek pricing and model limits — documented alias mapping, context, output and pricing.
- DeepSeek-V4-Flash-0731 model card — official checkpoint metadata, deployment example and MIT scope.
- DeepSeek V4 preview announcement and technical report — historical architecture context only.
Independent and hardware sources
- Artificial Analysis: Kimi K3 and DeepSeek V4 Flash — dated Intelligence Index entries.
- Artificial Analysis Kimi providers and DeepSeek providers — dated provider coverage and available performance fields.
- Apple Mac Studio technical specifications, Apple’s M3 Ultra announcement and current US configurator — memory configurations and bandwidth.
- NVIDIA DGX Spark hardware overview and system overview — memory, bandwidth and generic parameter envelope.
- DwarfStar/ds4 — community quantization guidance and Flash-family speed reports.
- MiaAI-Lab two-DGX-Spark 0731 recipe and pinned result artifact — version-pinned community two-Spark evidence.