Model System

Toolkit flagships.tk_ breadth.One router.

Toolkit is a routed AI system: Toolkit-owned SFT lanes for the flagship experience, plus supplemental tk_ routes that expose named models through one API key, one wallet, and token-level pricing.

Naming Rule

PrefixRailEconomicsWhat Users Should Expect
toolkit-Toolkit SFT railToolkit-priced product cardFirst-party behavior, near real-time SFT where applicable, and hardened free-rail throttles.
tk_Direct provider railProvider token card; platform economics live outside the cardNamed supplemental models for frontier, coding, reasoning, and long-context work.
toolkoder-Local SFT model railRuns on developer hardware; no hosted token cardDistilled local models for cheap bounded work; worker-first by default, local DAD capable for self-hosted flows.

Local ToolKoder SFT Models

ToolKoder local models are distilled SFT local models. The default Toolkode pairing is worker-first: a hosted Toolkit/provider captain plans, delegates, and reviews while local ToolKoder handles cheap, fast, bounded work. Self-hosted setups can also run ToolKoder as a local DAD/captain path when that tradeoff is what the developer wants.

Mac MLX lane
HardwareModelBaseRole
8GB Mactoolkoder-e4b-mlxGemma 4 E4BCurrent POC for cheap bounded local work; local DAD capable
16GB Mactoolkoder-9b-mlxQwen 9B targetNext local model for stronger coding turns and self-hosted DAD flows
24GB Mactoolkoder-35b-a3b-mlx35B MoE, ~3B activeStronger local model when MoE fits the Mac lane
36GB Mactoolkoder-27b-mlxQwen 27B denseQuality local model for larger bounded tasks and local DAD use
RTX NVFP4 lane
VRAMModelRole
8GBtoolkoder-e4b-nvfp4NVFP4 local model for worker-first or local DAD setups
10GB / 12GB / 16GBtoolkoder-9b-nvfp4NVFP4 local model for worker-first or local DAD setups
16GB / 24GBtoolkoder-35b-a3b-nvfp4NVFP4 local model for worker-first or local DAD setups
32GB / 48GBtoolkoder-27b-nvfp4NVFP4 local model for worker-first or local DAD setups

Canonical artifact paths use models/toolkoder-e4b and adapters/toolkoder-e4b-v152-exact-sequence, with the same pattern for 9b, 35b-a3b, and 27b.

Model Ladder

ModelParamsRoleInput / 1MCached / 1MOutput / 1M
toolkit-qwen3-4b4B denseNear real-time default chat for news, sports, stocks, and local context$0.08$0.008$0.34
toolkit-gemma4-E2B2.3B activeNear real-time ultra-cheap summaries and simple current-event turns$0.06$0.006$0.18
toolkit-gemma4-E4B4.5B activeNear real-time small SFT chat with stronger cheap answers$0.10$0.01$0.30
toolkit-qwen3.5-9b9B denseNear real-time mid-tier SFT chat for better reasoning at low cost$0.20$0.02$0.30
toolkit-qwen3.6-35B36B / 3B activeNear real-time premium SFT chat for richer real-world answers$0.496$0.0496$2.97
toolkit-qwen3.6-27B27B denseNear real-time premium quality SFT chat when output quality matters most$1.20$0.12$7.20
toolkit-glm-4.7-flashnot disclosedNear real-time always-on low-cost long-context SFT rail (200K)$0.12$0.012$0.80
tk_qwen37_maxnot disclosedFlagship captain lane for high-value planning and hard agent work$1.25$0.25$3.75
tk_qwen37_plusnot disclosedQwen specialist for multimodal reasoning, coding, and long-context work$0.40$1.16
tk_glm52open-weightReasoning, review, code, and 1M-context captain candidate$1.40$0.26$4.40
tk_kimi_k27_code1T / 32B active MoELong-horizon coding lead for repo work and large-context agent loops$0.95$0.19$4.00
tk_deepseek_v4_pro1.6T sparseFlagship captain lane for deep reasoning and code review$0.435$0.003625$0.87
tk_deepseek_v4_flash284B sparseBudget tk_ coding, reasoning, and agent turns$0.14$0.0028$0.28
tk_qwen36_plusnot disclosedOlder Qwen Plus fallback behind Qwen3.7 Plus$0.325$0.0325$1.95
tk_grok_buildnot disclosedBuilds, APIs, repo-aware implementation$1.00$0.20$2.00
tk_grok_code_fast_1not disclosedFast scoped coding worker and cheap code scout$0.20$0.02$1.50
tk_grok41_fastnot disclosedTool-calling, web research, code execution, and 2M-context scout
tk_groq_compound_minicompound routerCheap fast web-search, current-data, and computer-use scout$0.59$0.79
tk_cursor_composer_25Cursor workerPaid edit-loop implementation specialist$0.50$2.50
tk_gpt54_mininot disclosedVision, screenshots, documents, multimodal tasks$0.75$0.075$4.50
toolkit-voicevoice systemReal-time voice$0.10/min$0.10/min
toolkit-cam9B multimodalCamera / vision$0.10/min$0.10/min

Prices per 1M tokens unless noted. Cached input = repeat context and multi-turn history at 90% discount. Voice and camera billed per minute.

Hosted Captain Lanes

These are the primary hosted captain choices for planning, delegation, review, and hard agent work. Cheaper tk_ routes and local toolkoder- models can pair beneath them as sub-workers.

Captain LaneFamilyBest UseContextInput / 1MCached / 1MOutput / 1M
tk_qwen37_maxQwen 3.7 Max via AlibabaHigh-value planning, delegation, and hard general agent work1M$1.25$0.25$3.75
tk_glm52GLM 5.2Reasoning, review, long-horizon code, and tool orchestration1M$1.40$0.26$4.40
tk_kimi_k27_codeKimi K2.7 CodeLong-horizon coding, repo work, and high-context agent loops262K$0.95$0.19$4.00
tk_qwen37_plusQwen 3.7 Plus via AlibabaQwen specialist backup for coding, multimodal reasoning, and 1M-context work1M$0.40$1.16
tk_deepseek_v4_proDeepSeek V4 ProDeep reasoning and code review when value beats raw speed1M$0.435$0.003625$0.87

Captain Telemetry Matrix

Captain-facing model resolution matrix. toolkit- rows are near real-time SFT products for the current-data lane; tk_ rows are direct or provider-rate rails and are flagged as captain, sub-worker, or both. The tk_ order is sorted by a CFO utility score using Artificial Analysis intelligence, price, speed, latency, context, and DeepSWE coding-agent performance where Datacurve has an exact row.

External telemetry sources: Artificial Analysis for independent model intelligence, price, speed, latency, and context measurements; DeepSWE by Datacurve for long-horizon coding-agent pass rate, task cost, wall-clock time, and output-token aggregates. DeepSWE tasks and benchmark artifacts are never used for Toolkit training.

RankModelAgent RoleAA MatchAA ScoreAA BlendDeepSWE P@1DeepSWE P@4SWE $SWE TimeSWE OutTok/sContextSourceSignal
SFTtoolkit-qwen3-4b
Qwen3 4B
near real-timedirect12$0.132n/an/an/an/an/a95.132K servedAANear real-time default SFT chat; AA score estimated.
SFTtoolkit-gemma4-E2B
Gemma 4 E2B
near real-timedirect15n/an/an/an/an/an/an/a128KAANear real-time ultra-cheap edge SFT; AA lists quality, not hosted speed.
SFTtoolkit-gemma4-E4B
Gemma 4 E4B
near real-timedirect19n/an/an/an/an/an/an/a128KAANear real-time small edge SFT; AA lists quality, not hosted speed.
SFTtoolkit-qwen3.5-9b
Qwen3.5 9B
near real-timedirect32$0.105n/an/an/an/an/a68.4262KAANear real-time mid SFT chat; 2x base-model token card.
SFTtoolkit-qwen3.6-35B
Qwen3.6 35B A3B
near real-timedirect43$0.372n/an/an/an/an/a174.2262KAANear real-time premium speed/quality SFT chat balance.
SFTtoolkit-qwen3.6-27B
Qwen3.6 27B
near real-timedirect46$0.9n/an/an/an/an/a56.8262KAANear real-time premium quality chat; expensive output.
SFTtoolkit-glm-4.7-flash
GLM-4.7-Flash
near real-timedirect30$0.103n/an/an/an/an/a80.3200K servedAANear real-time always-on long-context Toolkit SFT rail.
#1tk_mimo_v25_pro
MiMo-V2.5-Pro
bothdirect54$0.17619.5%45.1%$1.9927.6m49K501MAA / DeepSWECan captain bounded work, but usually best as a strong implementation worker.
#2tk_gpt54_mini
GPT-5.4 mini (xhigh)
bothdirect49$0.65224.3%46%$2.0832.6m135K160.9400KAA / DeepSWEStrong captain-worker hybrid for coding and multimodal work under the $5 output ceiling.
#3tk_kimi_k27_code
Kimi K2.7 Code
captaindirect42$0.731%n/a$2.82149 steps59K53.1262KAA / DeepSWECurrent long-horizon coding lead with exact DeepSWE row and official K2.7 coding upgrade evidence.
#4tk_glm52
GLM-5.2
captaindirect51n/a28.6%n/a$6.4725.2mn/a1011MAA / DeepSWEStrong new reasoning/review captain candidate with live AA Coding Agent evidence; canary before silent default promotion.
#5tk_glm51
GLM-5.1
bothdirect51$0.90217.5%38.9%$7.4634.8m49K61.9200KAA / DeepSWEFallback reasoning/review rail; GLM 5.2 supersedes it for new captain trials.
#6tk_deepseek_v4_pro
DeepSeek V4 Pro (Max)
captaindirect52$0.1767.5%18.6%$4.2237m50K47.61MAA / DeepSWEFlagship captain lane for deep reasoning and code review.
#7tk_mimo_v25
MiMo-V2.5
sub-workerdirect49$0.058n/an/an/an/an/a92.11MAAStrong cheap 1M-context worker rail.
#8tk_qwen37_max
Qwen3.7 Max via Alibaba
captaindirect57$1.43n/an/an/an/an/a189.21MAAFlagship planning captain. Qwen routes resolve through Alibaba/DashScope.
#9tk_deepseek_v4_flash
DeepSeek V4 Flash (Max)
sub-workerdirect47$0.058n/an/an/an/an/a106.51MAABest cheap tk_ value for bounded coding/reasoning work.
#10tk_qwen37_plus
Qwen3.7 Plus via Alibaba
bothdirect39$0.2519.2%n/a$6.2310.6mn/a50.31MAA / DeepSWEAlibaba Qwen specialist for coding, multimodal reasoning, and long-context work; not a default replacement for Max.
#11tk_minimax_m27
MiniMax-M2.7
sub-workerdirect50$0.2220.2%0.9%$0.734.9m60K101.2205KAA / DeepSWEFast worker rail; poor DeepSWE coding-agent result.
#12tk_minimax_m25
MiniMax-M2.5
sub-workerdirect42$0.288n/an/an/an/an/a197.7205KAAFast, cheaper productivity worker rail.
#13tk_cursor_composer_25
Composer 2.5 / Cursor CLI
sub-workerproxy51.8n/a15.9%n/a$0.0859.7mn/an/aCursorAA / DeepSWEPaid Cursor implementation specialist. Treat as a guided worker/harness, not a normal Toolkit model endpoint.
#14tk_qwen36_plus
Qwen3.6 Plus
bothdirect50$0.4352.7%9.7%$4.2534.9m67K53.21MAA / DeepSWEOlder Qwen Plus fallback; prefer Alibaba Qwen3.7 Plus for new specialist work.
#15tk_glm5
GLM-5
bothdirect50$0.66n/an/an/an/an/a68200KAAGeneral reasoning rail; GLM 5.2 is the preferred new reasoning candidate.
#16tk_mimo_v2_pro
MiMo-V2-Pro
sub-workerdirect49$1.2n/an/an/an/an/a53.51MAAUse newer MiMo 2.5 rails first.
#17tk_kimi_k25
Kimi K2.5
bothdirect47$0.556n/an/an/an/an/a37.3256KAAOlder Kimi rail; K2.7 Code is the preferred Kimi coding route.
#18tk_cerebras_gptoss_120b_moe
gpt-oss-120b (high)
sub-workerdirect33$0.195n/an/an/an/an/a323.7131KAAVery fast MoE utility worker rail.
#19tk_glm47_flash
GLM-4.7-Flash
sub-workerdirect30$0.103n/an/an/an/an/a80.3131K/200KAACheap fallback GLM worker route.
#20tk_glm_turbo
GLM-5-Turbo
sub-workerdirect47n/an/an/an/an/an/an/a200KAAFast GLM worker rail; hosted telemetry incomplete.
#21tk_grok_code_fast_1
Grok Code Fast 1
sub-workerdirect21.6n/an/an/an/an/an/an/a256KAAFast scoped coding worker only; do not promote to captain from self-reported SWE data alone.
#22tk_grok_build
not listed
bothunavailablen/an/a13.1%29.2%$6.644.4m52Kn/a256KAA / DeepSWEBuild-oriented captain-worker hybrid for backend/API/repo-aware implementation; DeepSWE measured.
#23tk_grok41_fast
Grok 4.1 Fast
sub-workerunavailablen/an/an/an/an/an/an/an/a2MAATool-calling, agentic search, code execution, and long-context research specialist.
#24tk_groq_compound_mini
not listed
sub-workerunavailablen/an/an/an/an/an/an/an/a131KAACheap fast web-search, current-data, and computer-use scout; return evidence packets for stronger captains to verify.

Snapshot: Artificial Analysis leaderboard observed 2026-06-19; DeepSWE aggregate leaderboard generated 2026-06-19 and observed 2026-06-19. AA blended price and DeepSWE task cost are benchmark telemetry, not the Toolkit customer token card. API usage should read the same telemetry from /v1/models.

Cheap Model First, Big Brain If Needed

Toolkit routes requests based on task type and difficulty, not just model size. The router attempts the lowest-cost capable model first, then escalates if necessary.

Task TypeDefault Model
Simple chat, summaries, extractiontoolkit-qwen3-4b
Very cheap current-event turnstoolkit-gemma4-E2B / toolkit-gemma4-E4B
Premium SFT chattoolkit-qwen3.6-35B / toolkit-qwen3.6-27B
Long-context business and general worktoolkit-glm-4.7-flash
Captain planning, delegation, and reviewtk_qwen37_max / tk_glm52 / tk_kimi_k27_code / tk_deepseek_v4_pro
Cheap code, SQL, and endpoint helptk_deepseek_v4_flash
Long-horizon coding agentstk_kimi_k27_code / tk_gpt54_mini / tk_cursor_composer_25 / tk_mimo_v25_pro
Backend architecture, APIs, systemstk_grok_build / tk_kimi_k27_code / tk_gpt54_mini
Web search, browser scouting, computer usetk_groq_compound_mini / tk_grok41_fast
Fast scoped code editstk_grok_code_fast_1 / tk_cursor_composer_25
Deep reasoning and code reviewtk_glm52 / tk_deepseek_v4_pro / tk_mimo_v25_pro
Cheap bounded ToolKoder work on local hardwaretoolkoder-* local worker or local DAD
Images, documents, screenshotstk_gpt54_mini
Voicetoolkit-voice
Camera / screenshotstoolkit-cam

Freshness & Privacy

toolkit- models are Toolkit-owned SFT rails with product behavior and current-event freshness where listed. tk_ models are direct provider rails and keep their provider-rate economics.

No user conversations are used for model training.

Model Details

ModelBaseTrainingCadenceRoleSignals
toolkit-qwen3-4b
4B dense
Qwen3 4B + near real-time current-data SFTNear real-time current-data SFT railNear real-time refreshNear real-time conversational model tuned for fresh information: war, sports, stocks, world news, local context, and current events.Conversational responses / Current events / Local information / Fast inference / Persona awareness
Data: News, sports, markets, weather, jobs, dining, travel
Prioritizes recent public information and drops stale material from the SFT pipeline.
toolkit-voice
voice system
Toolkit voice stackVoice behavior SFTDaily voice updatesFull-duplex voice conversation for real-time assistant workflows.Full-duplex voice / Natural conversation / Speaker profiles / Real-time interaction / Sub-second knowledge lookup
Data: Voice conversation patterns + live briefing context
toolkit-cam
9B multimodal
MiniCPM-o 4.5Multimodal SFT railPlannedPhone camera and screenshot understanding.Phone camera input / Screenshot understanding / Visual reasoning / Multimodal knowledge
Camera route planned for visual assistant workflows.
toolkit-glm-4.7-flash
not disclosed
GLM-4.7-Flash + Toolkit SFTNear real-time Toolkit SFT workhorse railNear real-time refreshNear real-time long-context workhorse for business analysis and documents.200K context / Low cost to serve / Business logic / Structured reasoning
Data: Current-data briefing context: news, markets, weather, regional data
This stays first-party because GLM-4.7-Flash is cheap enough to keep available when provider-rate tk_ rails are throttled or unavailable.

Add-on Pricing

Token prices live in the Model Ladder above. This section only lists non-token add-ons so the page does not repeat the same model price matrix twice.

Images
$0.03
per image (standard)
Web Search
$0.005
per search call
Code Exec
$0.005
per execution

Training Cadence

CadenceModels
Near real-timetoolkit-* SFT rails carry the current-data lane: news, war context, sports, stocks, world events, and local information refreshed through the Toolkit data loop
Captaintk_qwen37_max, tk_glm52, tk_kimi_k27_code, and tk_deepseek_v4_pro are the flagship hosted captain lanes for planning, delegation, review, and hard agent work
Local SFTtoolkoder-* distilled SFT models for MLX and NVFP4; worker-first under a hosted captain, local DAD capable for self-hosted flows
Provider verifiedtk_ model rates are provider-rate cards; stale or unconfigured rails degrade before spend
Daily voice updatesvoice quality and safety refresh

Key Design Principles

1.
Role specialization

Each model has a specific job instead of trying to be a single general model.

2.
Routed system architecture

Requests are routed to the lowest-cost capable model and escalated only when needed.

3.
SFT behavior layer

Toolkit-owned rails carry product behavior and freshness without exposing internal serving details.

4.
Freshness training for SFT

The Toolkit SFT line is refreshed through the near real-time current-data loop.

5.
Separation of storage and training

Conversation memory is stored for context and user experience, not used as training data.

Toolkit is a routed AI system composed of first-party toolkit- SFT rails plus supplemental tk_ provider rails for breadth. Toolkit-owned lanes carry our product behavior and SFT cadence, while tk_ routes expose named models such as GLM 5.2, Kimi K2.7 Code, Alibaba Qwen 3.7 Max/Plus, Groq Compound Mini, Grok, Composer 2.5, MiMo, MiniMax, DeepSeek, and Cerebras GPT-OSS through one API key. The system always selects the lowest-cost capable route, escalating only when necessary.

Try Toolkit Free
View plansAPI docs