An evidence-based guide to Moonshot AI’s October 23, 2026 release. Research checked October 23, 2026. This guide distinguishes vendor claims, independent evaluations and Kingy’s separately published three-task 3D pilot. No additional model tests were run for this update.
Kimi K3.1 Pro deserves serious consideration for coding agents and complex knowledge work. Independent testing places it close to the strongest proprietary models, with particularly large improvements in tasks that require software tools and multiple steps. Its more difficult comparison is with inexpensive open-weight alternatives: a higher overall score does not automatically make it the most economical model for a particular job.
Artificial Analysis gives K3.1 Pro an Intelligence Index score of 55, placing it fourth among model releases in its launch analysis. Moonshot’s earlier Kimi K3 scored 44. That is a substantial improvement on this evaluation suite, although the index is neither an IQ score nor a percentage of human-level intelligence. Artificial Analysis launch evaluation.
Moonshot promises downloadable weights by November 4, 2026. Moonshot release announcement. Until the files arrive, a self-hosting decision remains conditional. Hosted availability and a future download should occupy different rows in a procurement plan.
The evidence supports a strong, specific conclusion: K3.1 Pro is a credible frontier competitor for agentic work. It does not establish that Kimi has become the best model across every kind of reasoning, scientific work, software engineering, or factual question.
What Kimi K3.1 actually refers to
The released flagship is Kimi K3.1 Pro, accompanied by Kimi K3.1 Pro Highspeed. “K3.1” names the series; it should not be used to imply that every possible member of that series has shipped or shares identical specifications. Moonshot lists both Pro variants in its public API model catalogue. Kimi API model list.
Kimi also names several surrounding products. A model generates responses and tool requests. Kimi Code supplies a development environment and agent workflow. A consumer application can add search, document processing, execution tools, account limits, and its own interface. These layers affect what a user can accomplish.
That separation matters when watching a demonstration. A model producing an editable 3D asset through code execution is doing something different from a model directly generating a 3D file as its native output. Likewise, an agent that fixes a repository depends on its shell access, available packages, test environment, instructions, and stopping rules.
The verified specifications
| Item | What the available documentation establishes |
|---|---|
| Developer | Moonshot AI, maker of Kimi |
| Released variants | Kimi K3.1 Pro; Kimi K3.1 Pro Highspeed |
| Advertised scale | “3T-class”; Artificial Analysis lists 2.8 trillion total parameters and 104 billion active parameters |
| Context limit | Up to 1,048,576 tokens in Kimi’s documentation; product-tier restrictions can apply |
| Inputs | Text and images; Kimi Code also documents video input |
| Native output | Text, including code and structured tool requests |
| Reasoning settings | low, high, max; default max; thinking stays enabled |
| Public API identifiers | kimi-k3.1-pro, kimi-k3.1-pro-highspeed |
| Kimi Code identifiers | k3.1-pro, k3.1-pro-highspeed |
| Generation budget | API guide documents max_completion_tokens defaulting to 131,072, configurable up to 1,048,576 |
| Downloadable weights | Announced for release by November 4; unavailable at this research cutoff |
| Final weight licence | Not verified at this research cutoff |
| Knowledge cutoff | Not verified at this research cutoff |
Sources: Kimi Code model configuration, API quickstart, Artificial Analysis model details, and the release announcement linked above.
OpenRouter now lists moonshotai/kimi-k3.1-pro, served by Moonshot AI, at $2 per million input tokens, $10 output and $0.20 cached input. This is an additional hosted access route; it does not establish downloadable weights or an independently operated checkpoint. OpenRouter provider listing.
The generation setting is a documented request limit, not evidence that a user can combine a million-token prompt with a further million-token answer. Input, intermediate work, output, and the endpoint’s context accounting still constrain an actual request. Its enforcement was not independently tested for this article.
The modality descriptions also need precision. Artificial Analysis identifies its evaluated endpoint as text-and-image input with text output. Kimi Code documents video support at the product level. Neither statement establishes native audio generation, image generation, or video generation by the language model itself. A tool-enabled product can offer capabilities beyond the model’s native input/output interface.
Finally, a context window is capacity. It says how much material can be presented, not how reliably a model will use every part of that material. The long-context results later in this article are more useful evidence of retrieval and reasoning quality than the advertised window alone.
Architecture: A large model with sparse computation
K3.1 Pro extends K3 through substantially greater post-training, including reinforcement learning across hundreds of domains. Moonshot technical announcement.
The detailed published architecture belongs to Kimi K3, its predecessor. That model is a mixture of experts with 93 layers, 896 routed experts, 16 selected experts per token, and two shared experts. Its attention stack combines 69 Kimi Delta Attention layers with 24 gated Multi-head Latent Attention layers. The model card also specifies Attention Residuals, the MoonViT-V2 vision encoder, and quantization-aware training using MXFP4 weights and MXFP8 activations. These are architectural background for the K3 lineage, not a substitute for inspecting the final K3.1 weight release. Kimi K3 model card.
Mixture-of-experts architecture helps explain the difference between total and active parameters. The model contains a large collection of learned components, but a token passes through only a selected portion of that collection. Using the reported counts, 104 billion is approximately 3.7% of 2.8 trillion.
That calculation does not mean the model needs only 3.7% of the storage. The remaining experts still exist and must be available when routing selects them. Nor does the fraction translate directly into a speed advantage: communication between accelerators, memory movement, attention, routing, and serving software all contribute to execution time.
Post-training also deserves attention. Training an agent to choose tools, recover from failed commands, inspect intermediate output, and continue a long task can improve completed-work scores substantially. Those improvements need not produce an equally large gain on a closed-book science question. The independent results show exactly why those capabilities should be measured separately.
Where K3.1 Pro sits against frontier and open-weight models
The following snapshot uses Artificial Analysis’ Intelligence Index v4.3.2 and its estimated cost per index task. It selects one configuration per listed model. “Fallback” is retained where the evaluator includes it in the configuration name.
| Model and evaluated setting | Weight availability | AA Index | USD per AA task |
|---|---|---|---|
| Claude Opus 5.5, max + fallback | Closed | 58 | $5.98 |
| Claude Sonnet 5.5, max + fallback | Closed | 56 | $5.46 |
| Muse Wonder Full, max | Closed | 56 | — |
| Kimi K3.1 Pro | Promised | 55 | $2.11 |
| Claude Fable 5.1, max + fallback | Closed | 53 | $7.63 |
| GPT-6 Astra, max | Closed | 53 | $3.26 |
| Gemini 4 Argon, high | Closed | 53 | $1.99 |
| GPT-6.1 Sol, max | Closed | 52 | $0.72 |
| MiMo-V2.6-Pro | Available | 46 | $0.13 |
| DeepSeek V4.1 Flash 1017, max | Available | 45 | $0.26 |
| GLM-5.3, max | Available | 45 | $2.01 |
| Qwen3.8 Max, 0902 | Closed | 45 | $5.41 |
| Kimi K3, max | Available | 44 | $2.00 |
| Qwen3.8 2.4T A95B | Available | 40 | $2.16 |
| Muse Spark Open, xhigh | Available | 39 | $0.84 |
Source: Artificial Analysis leaderboard, October 23 snapshot. Scores are rounded. A dash means no cost is provided here, not free inference. “Available” describes weights, not unrestricted licensing.
The cost column is easy to misuse. It is a weighted estimate for this benchmark workload, incorporating the model’s token consumption and applicable token prices. It is not a quote for a coding session, a subscription, a completed client project, or a self-hosted deployment. Different tasks can change the ordering. Artificial Analysis cost methodology.
For procurement, the useful question is what happens when each candidate receives the same work and a realistic budget. A model that needs less human repair can justify a higher inference bill. A less expensive model can be the better choice when it already completes the task reliably. Both possibilities remain open after reading this table.
There is also a distinction between the leading model release and the leading configuration. Leaderboards can show several effort levels for the same model. Counting each row as a different competitor changes apparent rank without changing the underlying release.
The index combines ten evaluations across agent work, coding, general capabilities, and scientific reasoning. Its weighting is a choice about which capabilities matter. Changing that mix can change the preferred model, especially when a deployment has a narrow purpose. Intelligence Index methodology.
Independent evaluations: What improved, and what did not
Artificial Analysis’ current model data provides a more informative picture than a single aggregate. The table below compares the same evaluator’s Kimi K3 and K3.1 Pro records. Except for Elo, values are percentages; the hallucination metric has its own denominator, explained below.
| Independent evaluation | Kimi K3 | K3.1 Pro |
|---|---|---|
| Terminal-Bench 4.0 | 12.6 | 67.2 |
| AutomationBench-AA | 58.3 | 73.3 |
| GDPval-AA v2.1, Elo | 1,534 | 1,846 |
| Humanity’s Last Exam, text-only | 46.9 | 52.5 |
| SciCode | 59.5 | 60.8 |
| CritPt | 23.4 | 24.6 |
| MMMU-Pro | 80.5 | 83.9 |
| GDP.pdf, all-pass | 22.0 | 31.4 |
| AA-LCR v1.1 | 88.7 | 86.3 |
| AA-Omniscience hallucination rate ↓ | 53.2 | 43.9 |
Source: Artificial Analysis model data, including its K3 comparison data. The launch analysis corroborates the principal gains with rounded figures.
The standout movement is in agent execution. The academic and scientific rows change much less. That pattern makes a stronger case for testing Kimi on a repository repair or a document workflow than for assuming a similar improvement on every hard reasoning problem.
The long-context row is also worth retaining. It declines in this snapshot. Without uncertainty estimates and repeated controlled comparisons, that modest difference does not establish a general regression. It does, however, prevent a careful reader from treating “better at everything” as an empirical finding.
AA-LCR itself exercises document reasoning at roughly 100,000 input tokens. It does not certify reliable use of the entire advertised million-token window. Separately, AA’s site marks SciCode as under review, another reason to avoid building a sweeping conclusion on a small difference in that row. AA-LCR methodology, AA evaluation catalogue.
AutomationBench-AA measures average objective completion, with a task receiving zero when a guardrail is violated. Its headline score therefore allows partial completion; it is not the percentage of entire workflows completed successfully. AutomationBench-AA scoring.
GDPval’s Elo deserves a different reading from an accuracy percentage. AA evaluates professional deliverables through blind pairwise comparisons. Its value depends on comparative grading and the evaluated field. “1,846 Elo” does not mean 1,846 tasks completed, 84.6% success, or a fixed level of economic productivity. Keep the benchmark version and comparison date attached to the number. GDPval-AA methodology.
Arena adds a preliminary WebDev result
Arena now explicitly lists Kimi K3.1 Pro (Max). Its WebDev overall page, retrieved October 23 but labelled October 22, places Kimi eighth at 1,692 ±14, with 2,451 votes and a displayed rank spread of 7–13. The entry remains preliminary. This is a named leaderboard result; no inference about an anonymous model codename is needed.
| Selected WebDev configuration | Displayed rank | Arena score |
|---|---|---|
| Claude Opus 5.5, Max | 1 | 1,807 ±13 |
| GPT-6 Astra, Max | 2 | 1,774 ±10 |
| GPT-6.1 Sol, Max | 4 | 1,749 ±12 |
| Kimi K3.1 Pro, Max | 8 | 1,692 ±14, preliminary |
| DeepSeek-V4.1-Flash-20261017, Max | 14 | 1,665 ±13 |
| Kimi K3, Max | 17 | 1,655 ±6 |
Source: Arena WebDev overall. These are selected rows, not consecutive places. The HTML-only category separately lists Kimi sixth at 1,702 ±28, with 596 votes and a 3–13 rank spread. Arena HTML results.
WebDev Arena uses user preferences between generated applications. Its rating describes comparative preference in that setting; it does not measure the percentage of production projects delivered correctly. The preliminary label and rank spread matter more than a claim of an exact, settled position. Arena evaluation method, ranking method.
ARI Bench: A provisional autonomous ML-engineering result
ARI Bench asks an agent to build a small model from scratch within one hour on one GPU, then grades the submitted artifact on 100 hidden completion questions. Its score concerns this constrained engineering task. It should not be read as K3.1 Pro's accuracy on ordinary questions. ARI Bench methodology.
The October 23 report gives Kimi K3.1 Pro at max effort 20/100 hidden exact matches in 43 minutes 48 seconds. It used the direct Kimi Code subscription API, OpenCode 1.18.32 and a 128,000-token response allowance. The final artifact combined task-specific rules, a trained 2.9-million-parameter transformer and context matching. The publisher reports that the unchanged package loaded successfully; public validation was 25/25.
This is one valid seed of the three required for an official estimate. The report records no operator repair or assisted regrade. Its $3.84 API-equivalent estimate includes reasoning tokens once as output; actual subscription billing is unverified. An earlier attempt interrupted by faulty temperature telemetry is retained separately and excluded from that result and cost. The hidden score does not establish how much each component contributed. ARI Bench run report.
Why two Terminal-Bench scores can both be correct
Moonshot: 65.7%, Claude Code. Artificial Analysis: 67.2%, mini-swe-agent. These are separate runs with different agent software. Moonshot benchmark footnotes, Artificial Analysis Terminal-Bench evaluation.
Artificial Analysis evaluates all 66 Terminal-Bench 4.0 tasks and averages pass@1 over three repeats. Each task must pass its tests. Its same-method results put Sonnet 5.5 at 63.6% and Opus 5.5 at 59.6%, with their specified fallback configurations. That supports a leading result within AA’s setup. It does not authorize taking AA’s Kimi score, another lab’s Claude score, and a third lab’s GPT score and calling the result a controlled experiment. AA Terminal-Bench results and protocol.
Anthropic’s own Opus 5.5 report provides another useful caution: its 66.4% Terminal-Bench result uses xhigh effort, and it reports a standard error of ±2.6 percentage points. Its production safeguards can also route some categories of work to earlier models. An apparently precise decimal can therefore describe a system configuration with meaningful sampling uncertainty. Anthropic’s Opus 5.5 evaluation notes.
A coding benchmark measures the combination of model, prompts, tools, agent loop, runtime, and budget. The agent loop determines whether a failed command triggers a sensible repair, whether output is truncated, when context is compressed, and whether the model gets another attempt. These are substantive parts of the experiment.
For practical use, both kinds of evidence help. A common agent framework offers a more controlled comparison. A vendor’s preferred framework can reveal what its system accomplishes when configured for its strengths. Neither removes the need to test the environment a customer will actually deploy.
Moonshot’s wider benchmark claims, including the weaker results
This selection from Moonshot’s launch table includes both strong and weak comparisons. These are vendor-reported or vendor-compiled results, with mixed evaluation conditions.
| Evaluation | K3.1 Pro | Opus 5.5 | GPT-6 Astra | DeepSeek V4.1 Flash 1017 |
|---|---|---|---|---|
| Terminal-Bench Science 0.1 | 38.1 | 63.3 | 68.1 | 15.7* |
| DeepSWE v1.1 | 75.5 | 74.2 | 74.1 | 77.0 |
| SWE-Marathon v1.1 | 51.7 | 46.3 | 42.5 | 40.0 |
| ProgramBench, binary threshold | 31.5 | 65.0 | 50.0 | 29.1 |
| GPQA Diamond | 94.2 | 93.9 | 96.1 | 92.2 |
| HLE-Diamond | 39.4 | 60.9 | 65.8 | 21.6 |
| BenchCAD, tools, voxel IoU × 100 | 92.7 | 96.2 | 95.9 | 81.6 |
| OfficeQA Pro v2 | 64.4 | 62.2 | 54.4 | 40.4 |
Source: Moonshot’s table and footnotes. Science: older DeepSeek V4.1 Flash checkpoint. ProgramBench: ≥95% per-task threshold. Kimi BenchCAD: 1,000-case subset. Settings and fallbacks vary.
The table’s value is its range. Repository work, sustained engineering, hard academic reasoning, scientific workflows, document analysis, and CAD reconstruction are distinct demands. Combining them into “coding intelligence” can conceal the very differences a user needs to understand.
For example, a repository repair usually starts with an existing structure and an observable failure. Building a program from a description can require decisions about behavior that are not fully specified. Scientific work adds domain knowledge, experimental assumptions, and interpretation. A system can be excellent at one of these and considerably less dependable at another.
Small gaps also need restraint. Without matched conditions and uncertainty estimates, a decimal-point lead is a reason for further testing, not proof of a durable advantage. Larger gaps deserve investigation, but they still need a precise account of what was measured.
Reasoning scores require the complete benchmark name
Moonshot’s HLE Full: 70.0 with tools; 51.1 without. Moonshot results.
Those labels describe different tasks. Search and code execution can help an agent retrieve information, check arithmetic, explore candidates, and verify a proposed answer. A closed-book evaluation measures a different combination of knowledge and reasoning.
Artificial Analysis uses 2,158 text-only HLE questions from the 2,500-question revision, with pass@1 scoring. Its score should not be directly substituted for a full multimodal score or a score obtained with tools. HLE-Diamond is another separately named evaluation. Artificial Analysis HLE protocol.
When comparing models, preserve four labels beside an HLE number: the dataset version or subset, tool access, sampling or attempt budget, and evaluator. Without them, two accurate numbers can produce an inaccurate conclusion.
A strong CAD score does not certify a manufacturable part
BenchCAD’s Vision2Code task asks for executable CAD code from rendered views. Its voxel intersection-over-union metric measures overlap between reconstructed and reference geometry. That is a geometric similarity measure, not a percentage of arbitrary customer projects completed successfully. BenchCAD scoring documentation.
A useful downstream test should therefore inspect more than the render. Does changing a dimension preserve the intended relationships? Are the units correct? Do clearances remain valid? Can another person edit the model? Does the exported geometry open in the target application?
For game assets and web experiences, the analogous checks are different: frame rate, asset size, collision behavior, visual consistency, accessibility, and whether the result survives a second change request. A polished demonstration can show possibility. It cannot establish a typical success rate without a representative set of attempts.
Kingy’s 3D pilot: Geometry and delivery gave different answers
Kingy’s separate October 23 pilot compared K3.1 Pro, Opus 5.5, Astra and Argon on three synthetic assignments: a Blender bench, CAD bracket and Three.js configurator. Each used high reasoning, a 16,000-token output cap and native-provider routing through Vercel AI Gateway.
| Model | Complete first-pass tasks | Reported charges, all three workflows |
|---|---|---|
| Kimi K3.1 Pro | 1/3 | $0.4423 |
| Claude Opus 5.5 | 1/3 | $2.1069 |
| GPT-6 Astra | 2/3 | $1.9076 |
| Gemini 4 Argon | 1/3 | $1.2148 |
All four passed the initial CAD task. Kimi’s bench lost its wood appearance on GLB export; its browser scene had mobile-framing and camera-reset defects. Two model repairs addressed those failures, and all three final Kimi revisions passed.
Charges include initial answers, repairs and final revision requests; they exclude setup, rendering and review. Argon’s final bench revision remains unverified after a delivery timeout. This was one run per model per task, reviewed by the Codex assistant without an independent human panel. It establishes limited, inspectable outcomes rather than a statistically reliable ranking. Kingy’s pilot, exact model IDs and downloadable evidence.
Hallucinations: An improvement with limits
The independent AA-Omniscience result is encouraging, but its “hallucination rate” has a specific definition. It is incorrect answers divided by incorrect, partial, and unattempted answers combined. Correct answers are excluded from that denominator. Its separate accuracy measure counts correct answers across all questions. AA-Omniscience definitions.
Consequently, the table’s hallucination percentage is not the fraction of all sentences that are false, the fraction of everyday responses containing an error, or the chance that a particular business report is wrong. Those interpretations require different measurements.
A factuality assessment should examine errors and abstentions together. Refusing every difficult question can suppress one kind of failure while making the system less useful. Answering every question can make it look productive while increasing unsupported claims. Useful evaluation records correct answers, incorrect answers, partial answers, and admissions of uncertainty.
For research and financial analysis, the output should make verification practical. A reader should be able to follow a factual claim to a source, distinguish a reported figure from a calculated figure, reproduce an important calculation, and identify the date governing the analysis. A model can help assemble that work; fluency alone cannot validate it.
The same applies to citations. A reference may exist without supporting the sentence attached to it. Testing citation accuracy requires checking the relationship between claim and source, including units, periods, qualifiers, and whether the source describes a forecast or an observed result.
Pricing: The headline rates and the actual bill
Moonshot publishes the following USD rates per million tokens. These are API rates, separate from consumer and coding subscriptions.
| Billing item | K3.1 Pro | K3.1 Pro Highspeed | Kimi K3 |
|---|---|---|---|
| Uncached input | $2.00 | $4.00 | $3.00 |
| Output | $10.00 | $20.00 | $15.00 |
| Cached input | $0.20 | $0.20 | $0.30 |
| Cache write, 5-minute TTL | $2.00 | $4.00 | $3.00 |
| Cache write, 1-hour TTL | $4.00 | $8.00 | $6.00 |
Source: Kimi API pricing, including the published pricing table embedded in that page. Taxes and separately billed services are outside these rates.
The regular model’s base input and output rates are one-third lower than K3’s. They are also half Opus 5.5’s $4 input and $20 output rates. That comparison concerns token prices; it does not establish a 50% saving on completed work. Kimi platform rates, Anthropic pricing.
Cache accounting can make the difference between a useful estimate and a badly inflated one. For Kimi’s Chat Completions and Responses APIs, cache reads, cache writes, and the uncached remainder are mutually exclusive portions of total input. Do not charge the same written token again as uncached input. The documentation specifies different usage-field accounting for the Messages interface. Kimi cache accounting.
For regular K3.1 Pro, a token-cost calculation is:
USD = (2 × uncached + 0.20 × cache_reads + 2 × five_minute_writes + 4 × one_hour_writes + 10 × billed_output) / 1,000,000
Each variable is a token count. For illustration, 100,000 input tokens billed at the regular or five-minute-write rate, plus 10,000 billed output tokens, cost $0.30. If 90,000 input tokens instead hit the cache and 10,000 are new at the regular rate, the same output budget produces $0.138 in token charges. These are arithmetic examples, not measured task bills.
Long agent sessions complicate the forecast because they may repeatedly include earlier messages, tool output, and reasoning-related state. A large prompt can be affordable when a stable prefix is reused; repeated misses can change the economics. The useful operating measurement is cost per accepted result, including retries and human repair time.
Artificial Analysis illustrates the distinction. Its launch article quotes approximately $1.98 per index task, while its live model page showed $2.11 during this research. The article uses the latter snapshot rather than treating the two figures as interchangeable. The evaluator also reports about 55,000 output tokens per index task, above K3’s approximately 48,000. Cheaper tokens can coexist with similar total workload cost. AA launch cost discussion, live model page.
One small coding test shows why effort changes the bill
Atomic Agent's developers tested one Python package with five bugs, two missing features and 32 hidden tests, repeating each configuration three times in Atomic Agent v0.6.9. Their 96/96 total counts repeated checks on one task.
| Configuration | Checks passed | Mean time | Reported USD/task |
|---|---|---|---|
| K3.1 Pro, low | 96/96 | 2.2 minutes | $0.16–$0.21 |
| K3.1 Pro, max | 96/96 | 4.9 minutes | $0.35–$0.46 |
| Opus 5.5 | 96/96 | 69 seconds | $0.37 |
Kimi used its direct API at temperature 1; Opus used OpenRouter pinned to Anthropic. The Opus effort setting is unspecified in this report. Kimi's cost ranges reflect an unresolved discrepancy: six runs reduced the balance by $2.02, versus $1.52 calculated from logged tokens. The authors could not allocate that difference between effort levels; Opus's figure is reported billing.
The result supports trying lower effort on similar tasks. All configurations passed, so it cannot establish a quality winner or predict difficult repository work. It is an agent developer's self-reported experiment, not a Kingy replication. Atomic Agent's test and limitations.
Highspeed and the limits of a speed multiplier
Moonshot describes Highspeed as having the same model capability with approximately two to three times the output speed. In Kimi Code it consumes approximately twice the quota. These are provider claims about the service; an independent Highspeed comparison was not established for this article. Kimi Code documentation.
Artificial Analysis’ October 23, 22:13 UTC snapshot shows about 77.1 output tokens per second for the regular endpoint, compared with 73.5 in the earlier launch snapshot. Its Intelligence Index remains 55 and estimated index-task cost remains $2.11. These changing endpoint measurements do not establish a controlled speed improvement or guarantee performance for other prompts, regions or context lengths. Artificial Analysis performance data.
End-to-end work also includes time spent outside generation. Consider an illustrative task that spends half its time generating tokens and half waiting for builds, browser actions, or external tools. Even a threefold improvement in generation speed cuts total time only from 100 units to roughly 67, a 1.5× overall speedup.
That is why a team should measure time to an accepted result alongside tokens per second. More rapid generation is valuable, but it cannot make a slow test suite, a long download, or a blocked external service finish faster.
Open weights, open source, and realistic self-hosting
Kimi’s intended weight release matters because it could let other operators serve, study, adapt, and independently evaluate the model. The exact permissions will depend on the released licence and accompanying materials. “Open-weight” should not be silently expanded into “fully open-source.”
The Open Source Initiative’s definition addresses rights to use, study, modify, and share, together with access to parameters, relevant code, and sufficient information about training data. A downloadable checkpoint alone does not establish all of those conditions. OSI Open Source AI Definition.
The existing alternatives already illustrate the range:
| Published weight release | Licence identified in its repository |
|---|---|
| DeepSeek V4.1 Flash 1017 | MIT |
| MiMo-V2.6-Pro-RL | MIT |
| GLM-5.3 | GLM-5.3 licence |
| Qwen3.8 2.4T A95B | Qwen3.8-Max licence |
| Muse Spark Open | Muse Model License |
Repository sources: DeepSeek, Xiaomi MiMo, Z.ai, Qwen, and Meta. These are licence identities, not a legal assessment of a proposed use.
Names require particular care. The hosted Qwen3.8 Max 0902 entry and the published Qwen3.8 2.4T A95B checkpoint are separately evaluated releases. Muse Wonder Full and Muse Spark Open are also different models. A laboratory offering some open weights does not make every model it sells an open-weight model.
Hardware is the other constraint. Using 2.8 trillion parameters, the arithmetic storage floor is approximately:
| Representation | Weight values alone, decimal units |
|---|---|
| 16 bits per parameter | 5.6 TB |
| 8 bits per parameter | 2.8 TB |
| 4 bits per parameter | 1.4 TB |
These are calculated lower bounds, not published K3.1 package sizes or RAM requirements. They exclude quantization metadata, higher-precision tensors, runtime buffers, context state, and serving overhead. They also say nothing about whether a particular quantization preserves quality.
The reported active-parameter count does not turn the full model into a typical desktop download. A serious deployment plan needs the final checkpoint format, supported inference software, accelerator memory, interconnect capacity, parallelism strategy, target concurrency, and long-context memory measurements.
Self-hosting can offer control over where prompts are processed and how the service operates. It also transfers responsibility for access control, updates, observability, reliability, and capacity planning to the operator. The economic comparison is therefore against the full cost of operating a service, not against an API price multiplied by a guessed number of tokens.
Security and deployment behavior
Moonshot reports filtering cyber-related training data to reduce offensive capability before the weight release. Moonshot’s security discussion. That intervention needs to be evaluated on its own terms; it cannot establish deployment safety by itself.
Capability evaluations and behavioral safety evaluations answer different questions. Completing a security task measures whether a system can do it. Resisting a malicious instruction, staying within authorization, protecting credentials, and avoiding unintended changes require separate tests.
For any tool-using model, the deployment boundary should be explicit. A repository file, retrieved document, or webpage may contain instructions that conflict with the user’s task. The surrounding agent needs to handle that content as evidence rather than authority. A benchmark lead does not remove this application-level responsibility.
There is also an unusually concrete integration risk in the Kimi Code documentation: a mistyped model ID can fall back to K2.8 Preview without thinking. The documentation recommends starting a new session when switching into K3.1 and keeping the reasoning configuration stable. A comparison performed through an incorrectly configured client may therefore evaluate a different model from the one the tester intended. Kimi Code routing notes.
Record the returned model identity, endpoint, client version, and reasoning setting in an evaluation log. Preserve the full assistant messages required by the API’s multi-turn protocol. Otherwise, an integration error can masquerade as a model-quality problem. Kimi API usage guidance.
How to choose between Kimi and its competitors
Against Claude, the evidence supports a split decision. Kimi leads AA’s terminal evaluation, while Opus leads its professional-deliverable comparison. The practical choice depends on whether the workload rewards terminal execution, deliverable quality, or some combination of the two. A broad claim that one model replaces the other would erase a useful distinction. AA terminal comparison, AA professional-work comparison.
Against GPT-6 Astra and GPT-6.1 Sol, keep capability and budget separate. Astra’s release emphasizes scientific and technical work, including research-level mathematics; those tasks deserve their own evaluation rather than a decision based only on the aggregate table. Sol supplies a different cost-performance option in the AA snapshot. A useful trial would include both the demanding research task and the routine work surrounding it. GPT-6 Astra technical release.
Against Gemini 4 Argon, the overall index hides a meaningful reversal: AA reports 77.5% for Argon on AutomationBench-AA, ahead of Kimi’s 73.3%. A team choosing a model for business-software workflows should investigate that result even though Kimi ranks higher on the composite. AA automation comparison.
Against DeepSeek, MiMo, and GLM, Kimi’s broad capability score is only one consideration. These competitors already have published weights. DeepSeek and MiMo use MIT licences, while GLM has model-specific terms. Existing deployments, adaptation work, serving expertise, and workload cost can outweigh a benchmark advantage. Replacing a working system should have a measured benefit. DeepSeek release, MiMo repository, GLM repository.
For complex coding and knowledge-work agents, K3.1 Pro belongs in a serious comparison. Give it an existing project, realistic tools, and an acceptance test that another person can inspect. Include at least one follow-up change, because maintaining a result can expose weaknesses that a single initial generation misses.
For scientific analysis and difficult closed-book reasoning, evaluate those tasks directly. A model chosen for terminal execution may not be the best choice for identifying an invalid experimental assumption or solving a specialized theoretical problem. Preserve the original problem statements and have qualified reviewers assess reasoning as well as final answers.
For high-volume routine work, include inexpensive alternatives before committing to a flagship. Classify the failures that matter: an extraction omission, a formatting error, an unsupported factual statement, or a tool action with the wrong arguments. Average quality can conceal an unacceptable failure category.
For 3D and design work, grade the editable asset and its behavior, not only an attractive screenshot. Ask the model to revise a dimension, reuse a component, repair an export, or meet a performance constraint. These tests expose whether the result is maintainable.
For private infrastructure, wait for an inspectable K3.1 release or evaluate a checkpoint that is already available. A procurement decision should reference an exact model revision and licence, not a product family or an announced future state.
What a convincing hands-on evaluation would add
The current evidence is sufficient to justify testing K3.1 Pro. A stronger operational recommendation needs a workload-specific evaluation. A practical pilot could use 30–50 representative tasks distributed across the work a team actually performs, while recognizing that such a pilot cannot measure rare failures reliably.
Before running it, define what counts as an accepted result. For code, use held-out tests and review the implementation. For document research, check claim support and calculation accuracy. For a workflow, inspect the final application state and unintended side effects. For design, evaluate both appearance and editability.
Keep the comparison reproducible:
- Pin model identifiers, endpoint providers, client versions, tools, and reasoning settings.
- Use the same starting files and task instructions wherever the interfaces allow it.
- Run both a common-budget comparison and a clearly labelled comparison using each system’s preferred configuration.
- Repeat tasks that show unstable outcomes; retain failed trajectories as well as successful ones.
- Measure token charges, tool costs, elapsed time, human interventions, and repair time.
- Report completion rates by task category and severity of failure, alongside the overall average.
Blind human review can reduce the tendency to prefer an output because of the model name. Automated checks can catch objective errors that polished prose obscures. Both are useful when their role and limitations are explicit.
At this cutoff, the remaining evidence gaps include the released K3.1 weights and licence, reproducible self-hosted performance, independent Highspeed measurements, representative million-token tests, and broader behavioral evaluations. No unverified Arena identity or rumored benchmark is used here.
For a team deciding what to deploy, the next useful result is a record of accepted work under its own constraints: what finished correctly, what needed repair, how long it took, and what it cost.
Update history
October 23, 2026, afternoon update: Added Arena’s named preliminary WebDev results, Kingy’s separately published 3D pilot, Atomic Agent’s effort-and-cost experiment, ARI Bench’s provisional run, OpenRouter availability and a refreshed AA throughput snapshot. The promised weight release and unverified final licence remain unchanged.
Trending on Kingy
Keep reading with the stories getting the most attention now.
The Kingy Brief
Get The Kingy Brief.
AI changes, original tests and one practical thing to try. Fridays at 09:00 Vancouver time.
Free · Double opt-in · Unsubscribe anytime
Sign up for The Kingy Brief on Beehiiv.
Signup help and newsletter schedule
Signup form provided by Beehiiv. After submitting, check your inbox for "Confirm your subscription to The Kingy Brief" and open its confirmation link. Check Spam or Promotions if you cannot find it.
Fridays at 09:00 Vancouver time: source-checked AI changes, original tests and one practical thing to try. We skip a week when there is not enough verified material. Free. Unsubscribe anytime.