Inkling 2 Small on Tinker Today: 594B/17B MoE, AA Index 31, Apache Weights Next Week
---
description: "Thinking Machines Inkling 2 Small is on Tinker now: 594B/17B MoE, text+image, variable effort, up to 1M context; closed weights with Apache 2.0 intended next week; Tinker $0.30/$0.03/$1.00 sample; AA ~31. Hosted-vs-wait checklist. No Tinker chat call."
---
> Attributed compile + offline HTTP preflight on my box
> Primary source: Introducing Inkling 2 Small (Nov 5, 2026) + Inkling 2 Small model card by Thinking Machines Lab.
> Access / pricing: Tinker Pricing (page stamp “as of October 19, 2026â€); Tinker models & pricing docs (model id thinkingmachines/Inkling-2-Small); Tinker Playground (redirects to auth — no login on this box); tinker-cookbook.
> Release policy: A Safe Path to Open Weights (Jul 31, 2026).
> Independent index scrape: Artificial Analysis — Inkling 2 Small (intelligenceIndex ≈ 30.63, isOpenWeights: false in our HTML/JSON scrape).
> Secondary only: The Decoder (AA 31 vs predecessor Inkling Small 13); MarkTechPost URL intermittently Cloudflare-challenged — unused for numbers. Local copies: /workspace/reputation/runs/36-inkling-2-small/.
> Numbers below from the announcement, model card, Tinker pricing/docs, or AA scrape are vendor / dashboard unless marked otherwise. This box did not authenticate to Tinker, call sampling/fine-tune APIs, or download weights. Offline preflight = public HTTP scrapes only.
---
Thinking Machines Lab’s second generation landed on 2026-11-05 as Inkling 2 Small: a multimodal MoE you can poke on Tinker today, with closed weights at launch and a company line that full Apache 2.0 weights should follow next week after external safety testing and red-teaming. For builders who care about agent loops, fine-tune surfaces, and “when do I self-host?â€, that is a sharper decision than another Daybreak-gated cyber demo. The useful checklist is: what actually shipped, how you reach it on Tinker (and what it costs), what the Apache cliff is promising vs guaranteeing, how to read the short vendor/AA score map, and when to wait for downloadable weights instead of burning playground tokens.
What shipped (card + announcement)
Label these as vendor card / announcement:
| Field | Value |
|---|---|
| Architecture | Sparse MoE decoder — 594B total / 17B active |
| Routing (card) | 48 layers; each token → 8 of 384 experts; hybrid local + global attention |
| Modalities | Text + image in → text out (no audio on this model) |
| Vision | 292M encoder (Perception Encoder lineage); 20×20 patches + 2×2 pooling |
| Context | Up to 1M tokens (card); Tinker priced rows today emphasize 64K and 256K |
| Effort | Variable effort (RL conditioned on effort + length penalty) |
| Numerics (card) | MXFP8 and NVFP4 |
| License at launch | Tinker Terms of Use (closed weights) |
| Intended Apache drop | Full weights under Apache 2.0 — company intent “next week†after external safety / red-team (not a drop we verified) |
Training color from the announcement (still vendor narrative, not our run): long pretrain on GB300 NVL72 in native NVFP4, then a single joint RL run (~1,075 steps, 17M+ rollouts, 2,048 GPUs, 9 days) mixing coding, knowledge work, vision, tool use, and safety. Early coding checkpoints were tested on OpenCode under the name Fledge-Alpha — useful signal that the lab optimized for real agent harnesses, not only unit-test rewards.
Access matrix (what this box actually opened)
| Route | Preflight | Note |
|---|---|---|
| Announcement + model card | HTTP 200 | Params, benches, limitations |
| Tinker Pricing + models docs | HTTP 200 | $/M + model ids |
| Tinker Playground URL | 302 → auth | No session / no chat on this box |
| Safe-path blog | HTTP 200 | Staged open-weights policy |
| Artificial Analysis model page | HTTP 200 (retry) | Independent index ~30.63; still not open-weights |
| OpenRouter /api/v1/models | HTTP 200 | thinkingmachines/inkling + inkling-small listed; no Inkling 2 Small id yet |
| MarkTechPost | Cloudflare challenge / flaky | Unused for numbers |
Practical pin strings from Tinker docs (as scraped):
thinkingmachines/Inkling-2-Small
thinkingmachines/Inkling-2-Small:peft:262144 # 256K context row
Cookbook path named on the card: thinking-machines-lab/tinker-cookbook plus the tml-renderers package for render-related workflows. Checkpoints from Tinker fine-tunes are described as downloadable when the base weights ship.
Tinker economics (hosted now)
Prefer the pricing page + docs over aggregators. Training & sampling table (USD per 1M tokens, page dated October 19, 2026):
| Context row | Prefill | Cached prefill | Sample | Train |
|---|---:|---:|---:|---:|
| Inkling 2 Small 64K | $0.30 | $0.03 | $1.00 | $1.50 |
| Inkling 2 Small 256K | $0.30 | $0.03 | $1.00 | $3.00 |
Serverless inference beta (same page; Inkling family only): Inkling 2 Small 256K at $0.30 input / $0.03 cached input / $1.00 output per 1M tokens. Storage: $0.10 / GB / month for checkpoints. MoE note on the page: priced by active parameters, not total size — which is why a 594B/17B active model can sit near the Inkling-Small dollar band on sample rates.
For agent loops, cached prefill at $0.03 vs $0.30 is the real budget knob once you keep long tool transcripts warm. We did not measure cache hit rates or tok/s here.
The Apache cliff (intent, not a receipt)
Announcement wording, paraphrased: Inkling 2 Small starts on Tinker with closed weights, following the lab’s safe path to open weights staging story (API / hosted fine-tune before irreversible public weights). They intend to release full weights under Apache 2.0 next week after external safety testing and red-teaming.
Treat that as company intent dated Nov 5, not a calendar we verified. Re-check next week: Hugging Face (or wherever they post), license file literally saying Apache-2.0, whether LoRA/full checkpoints from Tinker become portable, and whether OpenRouter / third-party hosts appear. Until then, “open-weights MoE†headlines are premature — AA’s scrape still had isOpenWeights: false / proprietary categorization on 2026-11-06.
Short score map (vendor + AA — not our eval)
Announcement prose claims an Artificial Analysis Intelligence Index score of 31 (vs prior Inkling models). The same post’s comparison matrix lists AA Index (v5.0) = 30.6. Our AA.com scrape for inkling-2-small showed intelligenceIndex ≈ 30.63. The Decoder (secondary) frames that as 31 vs Inkling Small 13 — cite only as aggregator color; we did not independently resolve the predecessor’s AA page in this preflight.
Tiny coding slice from the vendor announcement table (effort 0.99 / temperature 1.0 / 1M trajectory limit — their harness, not ours):
| Bench (vendor-reported) | Inkling 2 Small | DeepSeek-V4.1-Flash-1017 | Kimi K3 |
|---|---:|---:|---:|
| FrontierCode v2.0 | 25.4% | 28.5% | 17.3% |
| DeepSWE v1.1 | 72.1% | 77.0% | 68.5% |
| SWEBench Pro v2 (Hard) | 86.3% | – | 88.2% |
| Vibe Code Bench v1.1 | 81.6% | 87.7% | 85.0% |
Do not treat that matrix as a certificate. Use it to shortlist tasks for your golden eval once playground or API access is real. We invent no independent coding scores and we did not re-run any row.
Limitations they state (read before you demos)
Vendor limitations section, compressed: still not the most succinct communicator; can pick suboptimal tools in skill-heavy repos; no audio; chat UX is weaker (organization/presentation, slash overuse); can overreact on mental-health / safety-adjacent queries. Card also lists ordinary foundation-model risks (hallucination, uneven multilingual/domain coverage, cutoff). If your product is a polished consumer chat, assume extra formatting and refusal tuning work — even when coding-agent demos look strong.
Builder decision: Tinker now vs wait for Apache
Try Tinker now if you need agentic coding / vision-in-the-loop prototypes this week, want LoRA / custom RL on their stack before weights exist elsewhere, or you’re validating whether Fledge-Alpha-class behavior helps your OpenCode / harness loop. Pin thinkingmachines/Inkling-2-Small (or the :peft:262144 256K id), keep effort explicit, and budget cached-prefill math on long rollouts.
Wait for Apache weights if you need air-gap, weight-level inspection, non-Tinker serving, or license certainty before a production bet. “Next week†is the intent; verify the artifact + Apache text before you plan a cluster. Self-host hardware for a 594B MoE (even at 17B active) will still be a data-center story — this preflight did not find a consumer GGUF path because weights are not public yet.
Skip theater if you only wanted OpenRouter one-liner access today: our models API snapshot had prior Inkling ids, not Inkling 2 Small. Prefer Tinker until third-party routes appear.
What this box actually did
preflight 2026-11-06 ~09:22–09:28 America/Toronto
NO TINKER KEY / NO CHAT COMPLETION / NO WEIGHT DOWNLOAD
HTTP 200: announcement, model card, safe-path blog, tinker home,
tinker pricing, tinker docs models markdown, inkling product page,
AA models/search scrape + AA /models/inkling-2-small (retry),
OpenRouter /api/v1/models, The Decoder article, tinker-cookbook GH
HTTP 302→auth: tinker.thinkingmachines.ai/playground
HTTP 403 / Cloudflare challenge: MarkTechPost (first); UA path flaky — unused for numbers
HTTP 404: thinkingmachines.ai/news/a-safe-path-to-open-weights/ ;
thinkingmachines.ai/tinker/playground/
Parsed: 594B/17B; AA vendor 31 / matrix 30.6 / AA.com ~30.63;
model id thinkingmachines/Inkling-2-Small; Tinker $0.30/$0.03/$1.00
sample @64K; Apache “next week†intent; OR missing inkling-2
NOT tested: playground chat; fine-tune; serverless inference;
weight download; golden eval; latency/tok-s
Inkling 2 Small is a real Tinker-day release for agent builders: 594B/17B MoE, text+image, variable effort, up to 1M context on the card, hosted fine-tune/sampling priced today, and an Apache 2.0 next week intent that still needs a receipt. For most Toronto product teams on 2026-11-06, the rational move is playground + pin the model id + run your harness, treat AA ~31 as orientation not proof, and put a calendar reminder on the weight drop — not a cluster PO until the Apache file exists.
---
YongBo Yu — Toronto AI engineer (agent systems / LLM workflows). Site: yongbo-yu.vercel.app · GitHub: YongBoYu1
Sloth-Butler: Replies to the client messages you dread
This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
{% embed youtu.be %}
What I Built
Sloth-Butler handles the messages you dread replying to.
If you sell your time or your products, you know these messages. A client wants "just one more page" for free. A customer asks for 30% off. An invoice is three weeks late and the client wants another month. None of these are hard decisions. The hard part is the wording, so the reply sits in your inbox for days.
Sloth-Butler writes that reply in a few seconds: polite, firm, and with exactly one counter-offer, such as a paid option, a later date or a partial yes. It quotes your real rates, not numbers it made up.
What it handles:
- Situations: scope creep (free extra work), discount requests, late or unpaid invoices, pay-later or credit requests, last-minute favours or meetings, and "I'll get back to you". Each situation tells the model what a sensible counter-offer looks like.
- Your terms, enforced on the server. You save your rate, minimum project price, biggest discount you'll give, payment terms, late fee and next availability. They go with every request. The server then checks the reply in code: if it quotes a price below your minimum (including after a discount) or a discount above your limit, it regenerates once, and if that still fails it appends a line with your real terms. The model's numbers are never the only safeguard.
- Follow-ups. For late invoices and clients who went quiet, one button plans a friendly nudge, a firmer reminder and a final notice with your late fee, at +3, +7 and +14 days (editable). They land on a Follow-ups board with copy, voice note, mark-sent and mark-paid buttons.
- A protected log. When you send a reply, log the hours and money it protected. Totals add up on the page. It's stored only in your browser.
- 6 languages: English, Spanish, French, Portuguese, German and Nigerian Pidgin.
- Selectable currency, USD by default (EUR, GBP, CAD, AUD, INR, JPY and more).
- Voice. Play any reply in the ElevenLabs voice Alice, or download it as an mp3 voice note for chat apps.
An example. Situation: free extra work, tone firm, terms $60/hour and $900 minimum.
> Since you're already doing the logo, could you also put together a few social media banners? Should be quick for you.
> I can't add the banners to the logo project. I can quote them separately at $60 an hour if you'd like them this month.
There's no account and no database, and message text isn't logged.
Demo
Live: **sloth-butler.vercel.app** (guide and FAQ at /guide)
How to use it:
1. Open the site and press Write a reply.
2. Fill in Your terms once and save them.
3. Pick the situation and paste the message you received.
4. Add the amount or a deadline if it matters, pick a tone (soft, firm, maybe-later) and a language.
5. Press Write my reply. Copy it, play it, or download the mp3.
6. For a late invoice, press Plan follow-ups and work through the board.
7. After sending, log the hours and money it protected.
Always read the reply before sending. The model doesn't know your contract or calendar, so make sure the counter-offer is one you'll honour.
Further down the page, Compare with the untrained model runs the same message through plain Qwen3-8B and through Sloth-Butler side by side, next to the test results.
Code
{% github comzzy/sloth-butler %}
MIT licensed: dataset, training and eval scripts, raw eval outputs, and the Next.js app.
How I Built It
Tinker: the fine-tune
A prompt alone didn't reliably give me one short counter-offer that respects the user's numbers, in natural Pidgin. So I trained it in.
| Setting | Value |
|---|---|
| Base model | Qwen3-8B (thinking disabled) |
| Method | LoRA, rank 32, supervised fine-tuning on Tinker |
| Training examples | 1,101 across 6 languages |
| Held-out test set | 100 messages, never used in training |
| Schedule | 204 steps, 3 epochs, batch 16, learning rate 2e-4 decaying to 0 |
| Train loss | 2.90 at step 0, 0.44 at the last step |
The training loop is adapted from tinker-cookbook's supervised recipe. The app samples from the trained adapter through Tinker's OpenAI-compatible /completions endpoint, using the same prompt format and qwen3_disable_thinking template it was trained on, plus a line with the situation and the user's terms.
ElevenLabs: the voice
POST /api/voice turns a reply into speech with the ElevenLabs voice Alice and caches the mp3. A lot of client and customer conversations happen in WhatsApp-style chats, where a short voice note is the normal reply, so you can play it or download it and send it as is. Follow-ups have a voice-note button too.
Results: before vs after training
Base Qwen3-8B and the fine-tune answered the same 100 held-out messages with the same prompt, temperature 0.7 and seed 2026. A judge model (gpt-oss-120b) scored each reply 1 to 5 per criterion; a pass is 4 or more.
| Criterion | Before (Qwen3-8B) | After (Sloth-Butler) |
|---|---|---|
| Passes all five | 41% | 63% |
| Sendable | 84% | 93% |
| Firm but polite | 84% | 97% |
| Has a counter-offer | 74% | 90% |
| Fits the situation | 67% | 76% |
| Correct language | 90% | 94% |
| Pidgin sounds natural (10 replies) | 10% | 50% |
| Average words per reply | 51 | 34 |
| Blind side-by-side preference | | 56/100 for the fine-tune |
The clearest signals are the "passes all five" jump and the counter-offer gain. Replies also got about a third shorter, which matters when the goal is a message you can send without editing.
Stack
Next.js on Vercel, Qwen3-8B + LoRA trained and served on Tinker, ElevenLabs text-to-speech, gpt-oss-120b for data generation and judging, browser local storage for terms, follow-ups and the log.
Limits
- The model can still get numbers wrong. For example, it sometimes offers a discount when the job is already at your minimum price. The server-side guard catches that and corrects the reply, but it's a code check, not the model getting it right.
- The judge is a single LLM from the same family that wrote the training replies, so it may favour the fine-tune's style. No human rating yet.
- 100 test messages is small; a few points on one criterion is within noise.
- Pidgin went from 10% to 50% natural on just 10 replies, not checked by native speakers. Better, still not good.
- Test messages were generated the same way as the training data, so real inboxes may be harder.
What's next
Human ratings from freelancers and native Pidgin speakers, more real-world test messages, a second training round focused on Pidgin and on staying within the user's terms without the guard, and reminders for follow-ups that are due.
Why Does Open Innovation Matter?
Open weights are what made this possible. I could fix the behaviour I wanted by training a small adapter, not by asking a vendor for a feature. Nigerian Pidgin is spoken by tens of millions of people and the base model handled it poorly; with an open model, one person with a dataset can start closing that gap. The teacher and judge are open-weight too, so anyone can rerun the pipeline, check the numbers, or swap the base model. And an 8B model writing about 34 words costs very little per reply.
The Lazy Payoff
The time saved isn't in typing. It's the days a message sits unanswered because you don't know how to say no, or ask for your money, without sounding rude. With Sloth-Butler that's about ten seconds: paste, pick the situation, copy. Follow-ups for unpaid invoices are planned for you, and the log shows the hours and money you didn't give away.
Prize Categories
- Best Use of Tinker: the core of the project is a Qwen3-8B LoRA fine-tune trained and served on Tinker, with a before/after evaluation on 100 held-out messages.
- Best Use of ElevenLabs: every reply and follow-up can be played or downloaded as a voice note in the ElevenLabs voice Alice.
Credits
Qwen3-8B by the Qwen team. Tinker and tinker-cookbook by Thinking Machines Lab. gpt-oss-120b. ElevenLabs for the voice.
Mentions
Inkling 2 Small on Tinker Today: 594B/17B MoE, AA Index 31, Apache Weights Next Week --- description: "Thinking Machines Inkling 2 Small is on Tinker now: 594B/17B MoE, text+image, variable effort, up to 1M context; closed weights with Apache 2.0 intended next week; Tinker $0.30/$0.03/$1.00 sample; AA ~31. Hosted-vs-wait checklist. No Tinker chat call." --- > Attributed compile + offline HTTP preflight on my box > Primary source: Introducing Inkling 2 Small (Nov 5, 2026) + Inkling 2 Small model card by Thinking Machines Lab. > Access / pricing: Tinker Pricing (page stamp “as of October 19, 2026â€); Tinker models & pricing docs (model id thinkingmachines/Inkling-2-Small); Tinker Playground (redirects to auth — no login on this box); tinker-cookbook. > Release policy: A Safe Path to Open Weights (Jul 31, 2026). > Independent index scrape: Artificial Analysis — Inkling 2 Small (intelligenceIndex ≈ 30.63, isOpenWeights: false in our HTML/JSON scrape). > Secondary only: The Decoder (AA 31 vs predecessor Inkling Small 13); MarkTechPost URL intermittently Cloudflare-challenged — unused for numbers. Local copies: /workspace/reputation/runs/36-inkling-2-small/. > Numbers below from the announcement, model card, Tinker pricing/docs, or AA scrape are vendor / dashboard unless marked otherwise. This box did not authenticate to Tinker, call sampling/fine-tune APIs, or download weights. Offline preflight = public HTTP scrapes only. --- Thinking Machines Lab’s second generation landed on 2026-11-05 as Inkling 2 Small: a multimodal MoE you can poke on Tinker today, with closed weights at launch and a company line that full Apache 2.0 weights should follow next week after external safety testing and red-teaming. For builders who care about agent loops, fine-tune surfaces, and “when do I self-host?â€, that is a sharper decision than another Daybreak-gated cyber demo. The useful checklist is: what actually shipped, how you reach it on Tinker (and what it costs), what the Apache cliff is promising vs guaranteeing, how to read the short vendor/AA score map, and when to wait for downloadable weights instead of burning playground tokens. What shipped (card + announcement) Label these as vendor card / announcement: | Field | Value | |---|---| | Architecture | Sparse MoE decoder — 594B total / 17B active | | Routing (card) | 48 layers; each token → 8 of 384 experts; hybrid local + global attention | | Modalities | Text + image in → text out (no audio on this model) | | Vision | 292M encoder (Perception Encoder lineage); 20×20 patches + 2×2 pooling | | Context | Up to 1M tokens (card); Tinker priced rows today emphasize 64K and 256K | | Effort | Variable effort (RL conditioned on effort + length penalty) | | Numerics (card) | MXFP8 and NVFP4 | | License at launch | Tinker Terms of Use (closed weights) | | Intended Apache drop | Full weights under Apache 2.0 — company intent “next week†after external safety / red-team (not a drop we verified) | Training color from the announcement (still vendor narrative, not our run): long pretrain on GB300 NVL72 in native NVFP4, then a single joint RL run (~1,075 steps, 17M+ rollouts, 2,048 GPUs, 9 days) mixing coding, knowledge work, vision, tool use, and safety. Early coding checkpoints were tested on OpenCode under the name Fledge-Alpha — useful signal that the lab optimized for real agent harnesses, not only unit-test rewards. Access matrix (what this box actually opened) | Route | Preflight | Note | |---|---|---| | Announcement + model card | HTTP 200 | Params, benches, limitations | | Tinker Pricing + models docs | HTTP 200 | $/M + model ids | | Tinker Playground URL | 302 → auth | No session / no chat on this box | | Safe-path blog | HTTP 200 | Staged open-weights policy | | Artificial Analysis model page | HTTP 200 (retry) | Independent index ~30.63; still not open-weights | | OpenRouter /api/v1/models | HTTP 200 | thinkingmachines/inkling + inkling-small listed; no Inkling 2 Small id yet | | MarkTechPost | Cloudflare challenge / flaky | Unused for numbers | Practical pin strings from Tinker docs (as scraped): thinkingmachines/Inkling-2-Small thinkingmachines/Inkling-2-Small:peft:262144 # 256K context row Cookbook path named on the card: thinking-machines-lab/tinker-cookbook plus the tml-renderers package for render-related workflows. Checkpoints from Tinker fine-tunes are described as downloadable when the base weights ship. Tinker economics (hosted now) Prefer the pricing page + docs over aggregators. Training & sampling table (USD per 1M tokens, page dated October 19, 2026): | Context row | Prefill | Cached prefill | Sample | Train | |---|---:|---:|---:|---:| | Inkling 2 Small 64K | $0.30 | $0.03 | $1.00 | $1.50 | | Inkling 2 Small 256K | $0.30 | $0.03 | $1.00 | $3.00 | Serverless inference beta (same page; Inkling family only): Inkling 2 Small 256K at $0.30 input / $0.03 cached input / $1.00 output per 1M tokens. Storage: $0.10 / GB / month for checkpoints. MoE note on the page: priced by active parameters, not total size — which is why a 594B/17B active model can sit near the Inkling-Small dollar band on sample rates. For agent loops, cached prefill at $0.03 vs $0.30 is the real budget knob once you keep long tool transcripts warm. We did not measure cache hit rates or tok/s here. The Apache cliff (intent, not a receipt) Announcement wording, paraphrased: Inkling 2 Small starts on Tinker with closed weights, following the lab’s safe path to open weights staging story (API / hosted fine-tune before irreversible public weights). They intend to release full weights under Apache 2.0 next week after external safety testing and red-teaming. Treat that as company intent dated Nov 5, not a calendar we verified. Re-check next week: Hugging Face (or wherever they post), license file literally saying Apache-2.0, whether LoRA/full checkpoints from Tinker become portable, and whether OpenRouter / third-party hosts appear. Until then, “open-weights MoE†headlines are premature — AA’s scrape still had isOpenWeights: false / proprietary categorization on 2026-11-06. Short score map (vendor + AA — not our eval) Announcement prose claims an Artificial Analysis Intelligence Index score of 31 (vs prior Inkling models). The same post’s comparison matrix lists AA Index (v5.0) = 30.6. Our AA.com scrape for inkling-2-small showed intelligenceIndex ≈ 30.63. The Decoder (secondary) frames that as 31 vs Inkling Small 13 — cite only as aggregator color; we did not independently resolve the predecessor’s AA page in this preflight. Tiny coding slice from the vendor announcement table (effort 0.99 / temperature 1.0 / 1M trajectory limit — their harness, not ours): | Bench (vendor-reported) | Inkling 2 Small | DeepSeek-V4.1-Flash-1017 | Kimi K3 | |---|---:|---:|---:| | FrontierCode v2.0 | 25.4% | 28.5% | 17.3% | | DeepSWE v1.1 | 72.1% | 77.0% | 68.5% | | SWEBench Pro v2 (Hard) | 86.3% | – | 88.2% | | Vibe Code Bench v1.1 | 81.6% | 87.7% | 85.0% | Do not treat that matrix as a certificate. Use it to shortlist tasks for your golden eval once playground or API access is real. We invent no independent coding scores and we did not re-run any row. Limitations they state (read before you demos) Vendor limitations section, compressed: still not the most succinct communicator; can pick suboptimal tools in skill-heavy repos; no audio; chat UX is weaker (organization/presentation, slash overuse); can overreact on mental-health / safety-adjacent queries. Card also lists ordinary foundation-model risks (hallucination, uneven multilingual/domain coverage, cutoff). If your product is a polished consumer chat, assume extra formatting and refusal tuning work — even when coding-agent demos look strong. Builder decision: Tinker now vs wait for Apache Try Tinker now if you need agentic coding / vision-in-the-loop prototypes this week, want LoRA / custom RL on their stack before weights exist elsewhere, or you’re validating whether Fledge-Alpha-class behavior helps your OpenCode / harness loop. Pin thinkingmachines/Inkling-2-Small (or the :peft:262144 256K id), keep effort explicit, and budget cached-prefill math on long rollouts. Wait for Apache weights if you need air-gap, weight-level inspection, non-Tinker serving, or license certainty before a production bet. “Next week†is the intent; verify the artifact + Apache text before you plan a cluster. Self-host hardware for a 594B MoE (even at 17B active) will still be a data-center story — this preflight did not find a consumer GGUF path because weights are not public yet. Skip theater if you only wanted OpenRouter one-liner access today: our models API snapshot had prior Inkling ids, not Inkling 2 Small. Prefer Tinker until third-party routes appear. What this box actually did preflight 2026-11-06 ~09:22–09:28 America/Toronto NO TINKER KEY / NO CHAT COMPLETION / NO WEIGHT DOWNLOAD HTTP 200: announcement, model card, safe-path blog, tinker home, tinker pricing, tinker docs models markdown, inkling product page, AA models/search scrape + AA /models/inkling-2-small (retry), OpenRouter /api/v1/models, The Decoder article, tinker-cookbook GH HTTP 302→auth: tinker.thinkingmachines.ai/playground HTTP 403 / Cloudflare challenge: MarkTechPost (first); UA path flaky — unused for numbers HTTP 404: thinkingmachines.ai/news/a-safe-path-to-open-weights/ ; thinkingmachines.ai/tinker/playground/ Parsed: 594B/17B; AA vendor 31 / matrix 30.6 / AA.com ~30.63; model id thinkingmachines/Inkling-2-Small; Tinker $0.30/$0.03/$1.00 sample @64K; Apache “next week†intent; OR missing inkling-2 NOT tested: playground chat; fine-tune; serverless inference; weight download; golden eval; latency/tok-s Inkling 2 Small is a real Tinker-day release for agent builders: 594B/17B MoE, text+image, variable effort, up to 1M context on the card, hosted fine-tune/sampling priced today, and an Apache 2.0 next week intent that still needs a receipt. For most Toronto product teams on 2026-11-06, the rational move is playground + pin the model id + run your harness, treat AA ~31 as orientation not proof, and put a calendar reminder on the weight drop — not a cluster PO until the Apache file exists. --- YongBo Yu — Toronto AI engineer (agent systems / LLM workflows). Site: yongbo-yu.vercel.app · GitHub: YongBoYu1
Sloth-Butler: Replies to the client messages you dread This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady {% embed youtu.be %} What I Built Sloth-Butler handles the messages you dread replying to. If you sell your time or your products, you know these messages. A client wants "just one more page" for free. A customer asks for 30% off. An invoice is three weeks late and the client wants another month. None of these are hard decisions. The hard part is the wording, so the reply sits in your inbox for days. Sloth-Butler writes that reply in a few seconds: polite, firm, and with exactly one counter-offer, such as a paid option, a later date or a partial yes. It quotes your real rates, not numbers it made up. What it handles: - Situations: scope creep (free extra work), discount requests, late or unpaid invoices, pay-later or credit requests, last-minute favours or meetings, and "I'll get back to you". Each situation tells the model what a sensible counter-offer looks like. - Your terms, enforced on the server. You save your rate, minimum project price, biggest discount you'll give, payment terms, late fee and next availability. They go with every request. The server then checks the reply in code: if it quotes a price below your minimum (including after a discount) or a discount above your limit, it regenerates once, and if that still fails it appends a line with your real terms. The model's numbers are never the only safeguard. - Follow-ups. For late invoices and clients who went quiet, one button plans a friendly nudge, a firmer reminder and a final notice with your late fee, at +3, +7 and +14 days (editable). They land on a Follow-ups board with copy, voice note, mark-sent and mark-paid buttons. - A protected log. When you send a reply, log the hours and money it protected. Totals add up on the page. It's stored only in your browser. - 6 languages: English, Spanish, French, Portuguese, German and Nigerian Pidgin. - Selectable currency, USD by default (EUR, GBP, CAD, AUD, INR, JPY and more). - Voice. Play any reply in the ElevenLabs voice Alice, or download it as an mp3 voice note for chat apps. An example. Situation: free extra work, tone firm, terms $60/hour and $900 minimum. > Since you're already doing the logo, could you also put together a few social media banners? Should be quick for you. > I can't add the banners to the logo project. I can quote them separately at $60 an hour if you'd like them this month. There's no account and no database, and message text isn't logged. Demo Live: **sloth-butler.vercel.app** (guide and FAQ at /guide) How to use it: 1. Open the site and press Write a reply. 2. Fill in Your terms once and save them. 3. Pick the situation and paste the message you received. 4. Add the amount or a deadline if it matters, pick a tone (soft, firm, maybe-later) and a language. 5. Press Write my reply. Copy it, play it, or download the mp3. 6. For a late invoice, press Plan follow-ups and work through the board. 7. After sending, log the hours and money it protected. Always read the reply before sending. The model doesn't know your contract or calendar, so make sure the counter-offer is one you'll honour. Further down the page, Compare with the untrained model runs the same message through plain Qwen3-8B and through Sloth-Butler side by side, next to the test results. Code {% github comzzy/sloth-butler %} MIT licensed: dataset, training and eval scripts, raw eval outputs, and the Next.js app. How I Built It Tinker: the fine-tune A prompt alone didn't reliably give me one short counter-offer that respects the user's numbers, in natural Pidgin. So I trained it in. | Setting | Value | |---|---| | Base model | Qwen3-8B (thinking disabled) | | Method | LoRA, rank 32, supervised fine-tuning on Tinker | | Training examples | 1,101 across 6 languages | | Held-out test set | 100 messages, never used in training | | Schedule | 204 steps, 3 epochs, batch 16, learning rate 2e-4 decaying to 0 | | Train loss | 2.90 at step 0, 0.44 at the last step | The training loop is adapted from tinker-cookbook's supervised recipe. The app samples from the trained adapter through Tinker's OpenAI-compatible /completions endpoint, using the same prompt format and qwen3_disable_thinking template it was trained on, plus a line with the situation and the user's terms. ElevenLabs: the voice POST /api/voice turns a reply into speech with the ElevenLabs voice Alice and caches the mp3. A lot of client and customer conversations happen in WhatsApp-style chats, where a short voice note is the normal reply, so you can play it or download it and send it as is. Follow-ups have a voice-note button too. Results: before vs after training Base Qwen3-8B and the fine-tune answered the same 100 held-out messages with the same prompt, temperature 0.7 and seed 2026. A judge model (gpt-oss-120b) scored each reply 1 to 5 per criterion; a pass is 4 or more. | Criterion | Before (Qwen3-8B) | After (Sloth-Butler) | |---|---|---| | Passes all five | 41% | 63% | | Sendable | 84% | 93% | | Firm but polite | 84% | 97% | | Has a counter-offer | 74% | 90% | | Fits the situation | 67% | 76% | | Correct language | 90% | 94% | | Pidgin sounds natural (10 replies) | 10% | 50% | | Average words per reply | 51 | 34 | | Blind side-by-side preference | | 56/100 for the fine-tune | The clearest signals are the "passes all five" jump and the counter-offer gain. Replies also got about a third shorter, which matters when the goal is a message you can send without editing. Stack Next.js on Vercel, Qwen3-8B + LoRA trained and served on Tinker, ElevenLabs text-to-speech, gpt-oss-120b for data generation and judging, browser local storage for terms, follow-ups and the log. Limits - The model can still get numbers wrong. For example, it sometimes offers a discount when the job is already at your minimum price. The server-side guard catches that and corrects the reply, but it's a code check, not the model getting it right. - The judge is a single LLM from the same family that wrote the training replies, so it may favour the fine-tune's style. No human rating yet. - 100 test messages is small; a few points on one criterion is within noise. - Pidgin went from 10% to 50% natural on just 10 replies, not checked by native speakers. Better, still not good. - Test messages were generated the same way as the training data, so real inboxes may be harder. What's next Human ratings from freelancers and native Pidgin speakers, more real-world test messages, a second training round focused on Pidgin and on staying within the user's terms without the guard, and reminders for follow-ups that are due. Why Does Open Innovation Matter? Open weights are what made this possible. I could fix the behaviour I wanted by training a small adapter, not by asking a vendor for a feature. Nigerian Pidgin is spoken by tens of millions of people and the base model handled it poorly; with an open model, one person with a dataset can start closing that gap. The teacher and judge are open-weight too, so anyone can rerun the pipeline, check the numbers, or swap the base model. And an 8B model writing about 34 words costs very little per reply. The Lazy Payoff The time saved isn't in typing. It's the days a message sits unanswered because you don't know how to say no, or ask for your money, without sounding rude. With Sloth-Butler that's about ten seconds: paste, pick the situation, copy. Follow-ups for unpaid invoices are planned for you, and the log shows the hours and money you didn't give away. Prize Categories - Best Use of Tinker: the core of the project is a Qwen3-8B LoRA fine-tune trained and served on Tinker, with a before/after evaluation on 100 held-out messages. - Best Use of ElevenLabs: every reply and follow-up can be played or downloaded as a voice note in the ElevenLabs voice Alice. Credits Qwen3-8B by the Qwen team. Tinker and tinker-cookbook by Thinking Machines Lab. gpt-oss-120b. ElevenLabs for the voice.