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
| Prefix | Rail | Economics | What Users Should Expect |
|---|---|---|---|
| toolkit- | Toolkit SFT rail | Toolkit-priced product card | First-party behavior, near real-time SFT where applicable, and hardened free-rail throttles. |
| tk_ | Direct provider rail | Provider token card; platform economics live outside the card | Named supplemental models for frontier, coding, reasoning, and long-context work. |
| toolkoder- | Local SFT model rail | Runs on developer hardware; no hosted token card | Distilled 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.
| Hardware | Model | Base | Role |
|---|---|---|---|
| 8GB Mac | toolkoder-e4b-mlx | Gemma 4 E4B | Current POC for cheap bounded local work; local DAD capable |
| 16GB Mac | toolkoder-9b-mlx | Qwen 9B target | Next local model for stronger coding turns and self-hosted DAD flows |
| 24GB Mac | toolkoder-35b-a3b-mlx | 35B MoE, ~3B active | Stronger local model when MoE fits the Mac lane |
| 36GB Mac | toolkoder-27b-mlx | Qwen 27B dense | Quality local model for larger bounded tasks and local DAD use |
| VRAM | Model | Role |
|---|---|---|
| 8GB | toolkoder-e4b-nvfp4 | NVFP4 local model for worker-first or local DAD setups |
| 10GB / 12GB / 16GB | toolkoder-9b-nvfp4 | NVFP4 local model for worker-first or local DAD setups |
| 16GB / 24GB | toolkoder-35b-a3b-nvfp4 | NVFP4 local model for worker-first or local DAD setups |
| 32GB / 48GB | toolkoder-27b-nvfp4 | NVFP4 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
| Model | Params | Role | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|---|
| toolkit-qwen3-4b | 4B dense | Near real-time default chat for news, sports, stocks, and local context | $0.08 | $0.008 | $0.34 |
| toolkit-gemma4-E2B | 2.3B active | Near real-time ultra-cheap summaries and simple current-event turns | $0.06 | $0.006 | $0.18 |
| toolkit-gemma4-E4B | 4.5B active | Near real-time small SFT chat with stronger cheap answers | $0.10 | $0.01 | $0.30 |
| toolkit-qwen3.5-9b | 9B dense | Near real-time mid-tier SFT chat for better reasoning at low cost | $0.20 | $0.02 | $0.30 |
| toolkit-qwen3.6-35B | 36B / 3B active | Near real-time premium SFT chat for richer real-world answers | $0.496 | $0.0496 | $2.97 |
| toolkit-qwen3.6-27B | 27B dense | Near real-time premium quality SFT chat when output quality matters most | $1.20 | $0.12 | $7.20 |
| toolkit-glm-4.7-flash | not disclosed | Near real-time always-on low-cost long-context SFT rail (200K) | $0.12 | $0.012 | $0.80 |
| tk_qwen37_max | not disclosed | Flagship captain lane for high-value planning and hard agent work | $1.25 | $0.25 | $3.75 |
| tk_qwen37_plus | not disclosed | Qwen specialist for multimodal reasoning, coding, and long-context work | $0.40 | — | $1.16 |
| tk_glm52 | open-weight | Reasoning, review, code, and 1M-context captain candidate | $1.40 | $0.26 | $4.40 |
| tk_kimi_k27_code | 1T / 32B active MoE | Long-horizon coding lead for repo work and large-context agent loops | $0.95 | $0.19 | $4.00 |
| tk_deepseek_v4_pro | 1.6T sparse | Flagship captain lane for deep reasoning and code review | $0.435 | $0.003625 | $0.87 |
| tk_deepseek_v4_flash | 284B sparse | Budget tk_ coding, reasoning, and agent turns | $0.14 | $0.0028 | $0.28 |
| tk_qwen36_plus | not disclosed | Older Qwen Plus fallback behind Qwen3.7 Plus | $0.325 | $0.0325 | $1.95 |
| tk_grok_build | not disclosed | Builds, APIs, repo-aware implementation | $1.00 | $0.20 | $2.00 |
| tk_grok_code_fast_1 | not disclosed | Fast scoped coding worker and cheap code scout | $0.20 | $0.02 | $1.50 |
| tk_grok41_fast | not disclosed | Tool-calling, web research, code execution, and 2M-context scout | — | — | — |
| tk_groq_compound_mini | compound router | Cheap fast web-search, current-data, and computer-use scout | $0.59 | — | $0.79 |
| tk_cursor_composer_25 | Cursor worker | Paid edit-loop implementation specialist | $0.50 | — | $2.50 |
| tk_gpt54_mini | not disclosed | Vision, screenshots, documents, multimodal tasks | $0.75 | $0.075 | $4.50 |
| toolkit-voice | voice system | Real-time voice | $0.10/min | — | $0.10/min |
| toolkit-cam | 9B multimodal | Camera / 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 Lane | Family | Best Use | Context | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|---|---|
| tk_qwen37_max | Qwen 3.7 Max via Alibaba | High-value planning, delegation, and hard general agent work | 1M | $1.25 | $0.25 | $3.75 |
| tk_glm52 | GLM 5.2 | Reasoning, review, long-horizon code, and tool orchestration | 1M | $1.40 | $0.26 | $4.40 |
| tk_kimi_k27_code | Kimi K2.7 Code | Long-horizon coding, repo work, and high-context agent loops | 262K | $0.95 | $0.19 | $4.00 |
| tk_qwen37_plus | Qwen 3.7 Plus via Alibaba | Qwen specialist backup for coding, multimodal reasoning, and 1M-context work | 1M | $0.40 | — | $1.16 |
| tk_deepseek_v4_pro | DeepSeek V4 Pro | Deep reasoning and code review when value beats raw speed | 1M | $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.
| Rank | Model | Agent Role | AA Match | AA Score | AA Blend | DeepSWE P@1 | DeepSWE P@4 | SWE $ | SWE Time | SWE Out | Tok/s | Context | Source | Signal |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SFT | toolkit-qwen3-4b Qwen3 4B | near real-time | direct | 12 | $0.132 | n/a | n/a | n/a | n/a | n/a | 95.1 | 32K served | AA | Near real-time default SFT chat; AA score estimated. |
| SFT | toolkit-gemma4-E2B Gemma 4 E2B | near real-time | direct | 15 | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 128K | AA | Near real-time ultra-cheap edge SFT; AA lists quality, not hosted speed. |
| SFT | toolkit-gemma4-E4B Gemma 4 E4B | near real-time | direct | 19 | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 128K | AA | Near real-time small edge SFT; AA lists quality, not hosted speed. |
| SFT | toolkit-qwen3.5-9b Qwen3.5 9B | near real-time | direct | 32 | $0.105 | n/a | n/a | n/a | n/a | n/a | 68.4 | 262K | AA | Near real-time mid SFT chat; 2x base-model token card. |
| SFT | toolkit-qwen3.6-35B Qwen3.6 35B A3B | near real-time | direct | 43 | $0.372 | n/a | n/a | n/a | n/a | n/a | 174.2 | 262K | AA | Near real-time premium speed/quality SFT chat balance. |
| SFT | toolkit-qwen3.6-27B Qwen3.6 27B | near real-time | direct | 46 | $0.9 | n/a | n/a | n/a | n/a | n/a | 56.8 | 262K | AA | Near real-time premium quality chat; expensive output. |
| SFT | toolkit-glm-4.7-flash GLM-4.7-Flash | near real-time | direct | 30 | $0.103 | n/a | n/a | n/a | n/a | n/a | 80.3 | 200K served | AA | Near real-time always-on long-context Toolkit SFT rail. |
| #1 | tk_mimo_v25_pro MiMo-V2.5-Pro | both | direct | 54 | $0.176 | 19.5% | 45.1% | $1.99 | 27.6m | 49K | 50 | 1M | AA / DeepSWE | Can captain bounded work, but usually best as a strong implementation worker. |
| #2 | tk_gpt54_mini GPT-5.4 mini (xhigh) | both | direct | 49 | $0.652 | 24.3% | 46% | $2.08 | 32.6m | 135K | 160.9 | 400K | AA / DeepSWE | Strong captain-worker hybrid for coding and multimodal work under the $5 output ceiling. |
| #3 | tk_kimi_k27_code Kimi K2.7 Code | captain | direct | 42 | $0.7 | 31% | n/a | $2.82 | 149 steps | 59K | 53.1 | 262K | AA / DeepSWE | Current long-horizon coding lead with exact DeepSWE row and official K2.7 coding upgrade evidence. |
| #4 | tk_glm52 GLM-5.2 | captain | direct | 51 | n/a | 28.6% | n/a | $6.47 | 25.2m | n/a | 101 | 1M | AA / DeepSWE | Strong new reasoning/review captain candidate with live AA Coding Agent evidence; canary before silent default promotion. |
| #5 | tk_glm51 GLM-5.1 | both | direct | 51 | $0.902 | 17.5% | 38.9% | $7.46 | 34.8m | 49K | 61.9 | 200K | AA / DeepSWE | Fallback reasoning/review rail; GLM 5.2 supersedes it for new captain trials. |
| #6 | tk_deepseek_v4_pro DeepSeek V4 Pro (Max) | captain | direct | 52 | $0.176 | 7.5% | 18.6% | $4.22 | 37m | 50K | 47.6 | 1M | AA / DeepSWE | Flagship captain lane for deep reasoning and code review. |
| #7 | tk_mimo_v25 MiMo-V2.5 | sub-worker | direct | 49 | $0.058 | n/a | n/a | n/a | n/a | n/a | 92.1 | 1M | AA | Strong cheap 1M-context worker rail. |
| #8 | tk_qwen37_max Qwen3.7 Max via Alibaba | captain | direct | 57 | $1.43 | n/a | n/a | n/a | n/a | n/a | 189.2 | 1M | AA | Flagship planning captain. Qwen routes resolve through Alibaba/DashScope. |
| #9 | tk_deepseek_v4_flash DeepSeek V4 Flash (Max) | sub-worker | direct | 47 | $0.058 | n/a | n/a | n/a | n/a | n/a | 106.5 | 1M | AA | Best cheap tk_ value for bounded coding/reasoning work. |
| #10 | tk_qwen37_plus Qwen3.7 Plus via Alibaba | both | direct | 39 | $0.25 | 19.2% | n/a | $6.23 | 10.6m | n/a | 50.3 | 1M | AA / DeepSWE | Alibaba Qwen specialist for coding, multimodal reasoning, and long-context work; not a default replacement for Max. |
| #11 | tk_minimax_m27 MiniMax-M2.7 | sub-worker | direct | 50 | $0.222 | 0.2% | 0.9% | $0.7 | 34.9m | 60K | 101.2 | 205K | AA / DeepSWE | Fast worker rail; poor DeepSWE coding-agent result. |
| #12 | tk_minimax_m25 MiniMax-M2.5 | sub-worker | direct | 42 | $0.288 | n/a | n/a | n/a | n/a | n/a | 197.7 | 205K | AA | Fast, cheaper productivity worker rail. |
| #13 | tk_cursor_composer_25 Composer 2.5 / Cursor CLI | sub-worker | proxy | 51.8 | n/a | 15.9% | n/a | $0.085 | 9.7m | n/a | n/a | Cursor | AA / DeepSWE | Paid Cursor implementation specialist. Treat as a guided worker/harness, not a normal Toolkit model endpoint. |
| #14 | tk_qwen36_plus Qwen3.6 Plus | both | direct | 50 | $0.435 | 2.7% | 9.7% | $4.25 | 34.9m | 67K | 53.2 | 1M | AA / DeepSWE | Older Qwen Plus fallback; prefer Alibaba Qwen3.7 Plus for new specialist work. |
| #15 | tk_glm5 GLM-5 | both | direct | 50 | $0.66 | n/a | n/a | n/a | n/a | n/a | 68 | 200K | AA | General reasoning rail; GLM 5.2 is the preferred new reasoning candidate. |
| #16 | tk_mimo_v2_pro MiMo-V2-Pro | sub-worker | direct | 49 | $1.2 | n/a | n/a | n/a | n/a | n/a | 53.5 | 1M | AA | Use newer MiMo 2.5 rails first. |
| #17 | tk_kimi_k25 Kimi K2.5 | both | direct | 47 | $0.556 | n/a | n/a | n/a | n/a | n/a | 37.3 | 256K | AA | Older Kimi rail; K2.7 Code is the preferred Kimi coding route. |
| #18 | tk_cerebras_gptoss_120b_moe gpt-oss-120b (high) | sub-worker | direct | 33 | $0.195 | n/a | n/a | n/a | n/a | n/a | 323.7 | 131K | AA | Very fast MoE utility worker rail. |
| #19 | tk_glm47_flash GLM-4.7-Flash | sub-worker | direct | 30 | $0.103 | n/a | n/a | n/a | n/a | n/a | 80.3 | 131K/200K | AA | Cheap fallback GLM worker route. |
| #20 | tk_glm_turbo GLM-5-Turbo | sub-worker | direct | 47 | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 200K | AA | Fast GLM worker rail; hosted telemetry incomplete. |
| #21 | tk_grok_code_fast_1 Grok Code Fast 1 | sub-worker | direct | 21.6 | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 256K | AA | Fast scoped coding worker only; do not promote to captain from self-reported SWE data alone. |
| #22 | tk_grok_build not listed | both | unavailable | n/a | n/a | 13.1% | 29.2% | $6.6 | 44.4m | 52K | n/a | 256K | AA / DeepSWE | Build-oriented captain-worker hybrid for backend/API/repo-aware implementation; DeepSWE measured. |
| #23 | tk_grok41_fast Grok 4.1 Fast | sub-worker | unavailable | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 2M | AA | Tool-calling, agentic search, code execution, and long-context research specialist. |
| #24 | tk_groq_compound_mini not listed | sub-worker | unavailable | n/a | n/a | n/a | n/a | n/a | n/a | n/a | n/a | 131K | AA | Cheap 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 Type | Default Model |
|---|---|
| Simple chat, summaries, extraction | toolkit-qwen3-4b |
| Very cheap current-event turns | toolkit-gemma4-E2B / toolkit-gemma4-E4B |
| Premium SFT chat | toolkit-qwen3.6-35B / toolkit-qwen3.6-27B |
| Long-context business and general work | toolkit-glm-4.7-flash |
| Captain planning, delegation, and review | tk_qwen37_max / tk_glm52 / tk_kimi_k27_code / tk_deepseek_v4_pro |
| Cheap code, SQL, and endpoint help | tk_deepseek_v4_flash |
| Long-horizon coding agents | tk_kimi_k27_code / tk_gpt54_mini / tk_cursor_composer_25 / tk_mimo_v25_pro |
| Backend architecture, APIs, systems | tk_grok_build / tk_kimi_k27_code / tk_gpt54_mini |
| Web search, browser scouting, computer use | tk_groq_compound_mini / tk_grok41_fast |
| Fast scoped code edits | tk_grok_code_fast_1 / tk_cursor_composer_25 |
| Deep reasoning and code review | tk_glm52 / tk_deepseek_v4_pro / tk_mimo_v25_pro |
| Cheap bounded ToolKoder work on local hardware | toolkoder-* local worker or local DAD |
| Images, documents, screenshots | tk_gpt54_mini |
| Voice | toolkit-voice |
| Camera / screenshots | toolkit-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
| Model | Base | Training | Cadence | Role | Signals |
|---|---|---|---|---|---|
| toolkit-qwen3-4b 4B dense | Qwen3 4B + near real-time current-data SFT | Near real-time current-data SFT rail | Near real-time refresh | Near 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 stack | Voice behavior SFT | Daily voice updates | Full-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 4.5 | Multimodal SFT rail | Planned | Phone 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 SFT | Near real-time Toolkit SFT workhorse rail | Near real-time refresh | Near 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.
Training Cadence
| Cadence | Models |
|---|---|
| Near real-time | toolkit-* SFT rails carry the current-data lane: news, war context, sports, stocks, world events, and local information refreshed through the Toolkit data loop |
| Captain | tk_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 SFT | toolkoder-* distilled SFT models for MLX and NVFP4; worker-first under a hosted captain, local DAD capable for self-hosted flows |
| Provider verified | tk_ model rates are provider-rate cards; stale or unconfigured rails degrade before spend |
| Daily voice updates | voice quality and safety refresh |
Key Design Principles
Each model has a specific job instead of trying to be a single general model.
Requests are routed to the lowest-cost capable model and escalated only when needed.
Toolkit-owned rails carry product behavior and freshness without exposing internal serving details.
The Toolkit SFT line is refreshed through the near real-time current-data loop.
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.