A generated title card reading 'Inkling 2 Mini vs Qwen 3.8', subtitled 'the tier that actually competes is Flash', with three pills reading 'Mini: Index 20.1', 'Flash: Index 29' and 'Same cost per task', above two cards contrasting a '20GB file, runs offline' with a 'hosted endpoint, 1M context' and a note to escalate only when needed. The OrcaRouter logo sits bottom-right.
Guides & Insights

Inkling 2 Mini vs Qwen 3.8: Beat by the Cheaper Tier

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The comparison everyone runs is the wrong one, and it flatters the new model. Inkling 2 Mini, published by Thinking Machines Lab on 20 November 2026, scores 20.1 on the Artificial Analysis Intelligence Index v5.0.1 — level with Qwen3.8-27B at 20, at $0.59 per index task against $3.53. That is the head-to-head Thinking Machines put in its own benchmark table, and on cost it is not close. But Q​wen 3.8 is not one model. It is a generation of five, and the tier that actually competes with a 20GB downloadable checkpoint is Qwen3.8-Flash, which Artificial Analysis scores at 29 for $0.57 per index task — nine points ahead, at the same cost per finished answer. Mini wins the comparison the launch post printed and loses the one a buyer would run.

Sourcing, since it drives the verdict: the index scores, cost per task, latency and context figures are Artificial Analysis measurements read on 21 November 2026, and every OrcaRouter price is quoted from our own catalogue. The percentage tables are Thinking Machines' launch harness, vendor-run, flagged where they appear; A​libaba's own claims for the Q​wen tiers are quoted as A​libaba's and have not been separately audited here.

First, which Q​wen 3.8 you mean

Treating the family name as an opponent is the mistake that makes most comparisons of this kind useless. Q​wen 3.8 spans a hosted flagship at $2.00/$6.00 per million tokens, an open 2.4-trillion-parameter checkpoint under a custom licence, a 27B dense model under Apache 2.0, a Flash tier at $0.15/$0.47, and the Flash-Next preview whose open weights sit behind the hosted Flash product. The members differ by more than a factor of ten in token price and by nine points on the same index. A single "vs Q​wen 3.8" verdict is a verdict about a blur.

• Qwen3.8-27B — 27B dense, Apache 2.0, text, image and video in, 256K context, Index 20 at xhigh effort, $0.50/$3.00 list, $3.53 per index task, 34.5 minutes per task.

• Qwen3.8-Flash — the hosted production tier at $0.15/$0.47, serving the same weights as the Flash-Next preview; the routed endpoint advertises a 1M-token window and a 131K maximum output, while Artificial Analysis records 256K for the open Flash-Next weights. It measures that model at Index 29, 180B total and 6B active, $0.57 per index task, 33.3 minutes.

• Qwen3.8-Max — the closed flagship, Index 34, $2.00/$6.00, $10.02 per index task. Not a competitor for Mini on cost and not the model to compare against.

• Inkling 2 Mini — 26.4B total, 3.4B active, Apache 2.0, 1M context on the card, Index 20.1, $0.15/$0.50 on Tinker, $0.59 per index task, 9.0 minutes.

So the useful comparison has two halves. Against the 27B dense model the launch post picked, Mini is a genuine cost win at equal score. Against the Flash tier, which costs the same and runs circles around it on the index, Mini's argument has to be about something other than price and score — and it is: the file.

Screenshot of the Artificial Analysis model page for Qwen3.8-Flash-Next, showing it above average in intelligence and well priced for its size but notably slow per task and very verbose, a score of 29 on the Intelligence Index against a median of 23, 156M tokens generated during the index run against a 104M median, pricing of $0.15 per 1M input and $0.47 per 1M output tokens, $0.57 per index task, 33.3 minutes per task, an 85.77-second time to first token, 74 output tokens per second, and a listed 256K-token context window.

The 27B tie, and why it is real

Both models are measured by the same evaluator on the same revision of the same index, so 20.1 against 20 is a real comparison and not a coincidence of harnesses. It is also a tie at the resolution limit of the instrument. The index is a composite of nine evaluations, several on held-out sets, and models within a point of each other are not reliably ordered by it. What the tie establishes is that Mini clears the same bar as a 27B dense model while burning 3.4 billion active parameters instead of 27 billion.

The cost column is where that matters. At $0.59 per index task against $3.53, Mini is roughly six times cheaper to finish the same evaluation, and it does it in nine minutes against 34 and a half. Output speed is 280 tokens per second against 53. Time to first answer token is 9.72 seconds against 144. In an agent loop that runs thirty turns, those numbers compound into something a user can feel.

None of that is free, and the compensating cost is verbosity. Mini emitted 161M output tokens to complete the index run, against 128M for the Q​wen model and a peer median of 85.7M. Paying $0.50 per million output for 1.9 times the typical output is still cheap; budgeting a fixed number of tokens per task for a model that talks this much is not. A team that meters by tokens rather than by tasks will find the six-times advantage shrinks toward three.

Where Mini genuinely beats every Q​wen tier

The vendor's coding table is the part of the launch that makes the model interesting, and it is also the part nobody has checked. Every row below is Thinking Machines' own harness on its own checkpoint, and at least one of them was run with the company's agent scaffold rather than a neutral one.

• DeepSWE v1.1 — Inkling 2 Mini 48.2% vs Qwen3.8-27B 42.2%.

• SWE-bench Pro v2, Hard — 77.9% vs 73.8%.

• SWE-bench Multilingual — 73.7% vs 70.0%.

• SWE-bench Verified — 79.2% vs 77.8%.

• Vibe Code Bench v1.1 — 77.9% vs 64.9%.

• Terminal Bench v2.1 — 71.6% vs 73.0%, a loss; on Terminal Bench v4.0 both print 5.6%.

• BrowseComp — 79.0% vs 61.6%; DeepSearchQA 85.3% vs 79.7%; MCP Atlas 82.7% vs 77.3%.

• MMMU Pro — 74.9% vs 76.3%, another narrow loss; OmniDocBench v1.6 92.8% vs 91.5%.

The rows Mini wins are the agentic ones — repository-scale patching, long-horizon search, tool use. The rows it loses are the precise-recall ones. That is the architecture showing through: looping buys serial computation per token, and serial computation is what a long agent trajectory needs. It does not buy stored facts, and a knowledge-heavy benchmark notices.

The factuality gap between Mini and the family is starker than any of those rows. On AA-Omniscience, the vendor table prints −8.6 for Mini with a 63.4% non-hallucination rate, against −6.4 and 71.4% for Qwen3.8-27B in the same column of the same table. A negative score means the model is worse than declining to answer. Artificial Analysis does not publish a separate omniscience figure for Mini that I could find, so this row rests entirely on the vendor harness — and it is still the most important number in the comparison for anyone deploying without a retrieval layer.

Where Flash ends the argument

Now the tier that matters. Qwen3.8-Flash is hosted and priced at $0.15 in and $0.47 out — the same input rate Tinker charges to sample Mini, and three cents cheaper on output. Artificial Analysis scores the Flash-Next weights behind it at 29. Per index task it costs $0.57, against Mini's $0.59. Nine index points, for two cents less.

That is not a close call on capability or on price, and it reframes what Mini is for. Mini's advantages over Flash are not on the scoreboard at all: it is a 20GB file that runs with no network, on a MacBook or a DGX Spark, with no per-token meter and no provider in the loop. Flash's advantages are nine index points, a hosted endpoint, and a 1M-token window on the routed API with nothing to download. If your workload has a network connection, the Q​wen side wins on every measured axis. If it does not, nothing on the Q​wen side is available to you at all.

The speed comparison also favours Mini, and by more than the score gap suggests it should. Mini runs at 280 output tokens per second with a 9.72-second first token; Flash-Next measures 74 tokens per second with an 85.77-second first token and 33.3 minutes per index task. For an interactive agent, that is the difference between a tool that responds and one you wait on — and it is the one place where a 5.4B-effective-parameter local model has a structural advantage over a 180B hosted one.

The practical arrangement

Both Q​wen tiers are on OrcaRouter, which makes the escalation path unusually cheap to build. Qwen3.8-27B routes at $0.33 per million input and $2.40 per million output — below A​libaba's $0.50/$3.00 list, because the provider's price is passed through rather than marked up. Qwen3.8-Flash routes at $0.15/$0.47 on a 1M-token window. One key reaches both, with automatic failover and the routing DSL to send a request to the cheap tier first and escalate only when it fails a check.

That is the shape this comparison actually wants. Mini, if you adopt it, sits on your own hardware doing the agentic work it is good at and paying nothing per token. The Q​wen tiers sit behind a single endpoint for the requests Mini is bad at — the ones that need a fact recalled correctly, a long document read, or a hard scientific inference — and the failover means a provider outage degrades the answer instead of breaking it. We do not host Inkling 2 Mini and this piece will not pretend otherwise; what the router removes is the second contract and the second integration for the half of the workload that goes to Q​wen.

Screenshot of OrcaRouter's model page for Qwen3.8 Flash, showing the model id qwen/qwen3.8-flash, vision, tools, JSON and reasoning tags, a dated description of it as Alibaba's multimodal reasoning model for coding, agentic workflows and long-video analysis, an OpenAI-compatible /v1/chat/completions endpoint at api.orcarouter.ai, an input price of $0.15 and output price of $0.47 per 1M tokens, a 1M-token context window with a 131K maximum output, a p50 time to first token of 10.00 seconds, and an option to add it to a comparison.

What would change the verdict

Two things, and both are cheap to watch. The first is an independent coding run on Mini's public weights. If the 48.2% on DeepSWE and 77.9% on SWE-bench Pro survive a neutral harness, Mini's agentic case gets much stronger than a 20.1 index score implies, because the index underweights exactly the long-horizon work the model is built for. If they do not survive, the launch table was a vendor table and the model is a 20 at heart.

The second is whether Qwen3.8-Flash's nine-point advantage holds under effort controls. Its index figure is a single configuration, and Q​wen models move several points between effort settings — the 27B model scores 20 at xhigh and materially lower when thinking is dialled back. A Flash tier that loses nine points at low effort is a different competitor from one that keeps them, and nobody has published that curve for this comparison.

Until one of those lands, the honest summary is a split verdict. On measured capability per dollar in a hosted setting, Q​wen 3.8 wins twice — the 27B model ties Mini on score and loses on cost, the Flash tier beats it on both. On the things a hosted model cannot do at all — run offline, run on a laptop, run with no meter — Inkling 2 Mini has no Q​wen answer, and the 20GB file is the entire argument. Buy the download if you need it to run where there is no network. Buy the API if you do not.

A generated two-column scoreboard titled 'Inkling 2 Mini vs Qwen3.8-Flash — the scoreboard'. The Inkling 2 Mini column reads Intelligence Index 20.1, active parameters 3.4B, cost per index task $0.59, output speed 280 tokens per second, first token 9.72 seconds, and runs offline: yes, 20GB. The Qwen3.8-Flash column reads Intelligence Index 29, active parameters 6B, cost per index task $0.57, output speed 74 tokens per second, first token 85.77 seconds, and runs offline: no, hosted only. A footer states that both columns are per Artificial Analysis and coding rows are vendor-reported.