LFM3-1.7B-Base vs Qwen 3.8: Wrong Weight Class, Right Question

The honest answer is that LFM3-1.7B-Base does not compete with Qwen3.8-27B, and it was never meant to. Ask the question properly and three different Qwen models come back: the open-weight 27B vision-language model that Alibaba published in August 2026, the hosted Qwen3.8-Flash at $0.15 per million input tokens and $0.47 per million output, and Qwen3.8-Max at $2.00 and $6.00. Liquid's new 1.73B checkpoint is raw pre-trained material for a phone, a laptop or a fine-tuning run; Qwen's line is a family of finished products. The useful comparison is not who scores higher. It is which deployment envelope you are actually buying.

That framing also explains why this matchup has no serious coverage yet. There is nothing to benchmark against: one side has a vendor table with 27 rows, the other has no published evaluation at all. So this piece does the comparison in the only order that survives contact with the facts — what each side ships, what it costs to run, and what you are allowed to do with it.

Three Qwen 3.8s, three different purchases

Qwen3.8-27B is the open-weight option. Alibaba published it in August 2026 under Apache 2.0, and it has since accumulated more than 7.5 million downloads. It is a 27B dense causal model with a vision encoder, 64 layers in a 16 × (3 × Gated DeltaNet → FFN, 1 × Gated Attention → FFN) layout, and a 262,144-token native context that Alibaba says extends to 1,000,000. It understands images and video, it has thinking control, and it explicitly targets long-horizon agentic work.

Qwen3.8-Flash is the hosted tier most teams will actually price first. On OrcaRouter's catalogue it is a 1M-context model supporting text, image and video input, with a 131K output ceiling and rates of $0.15 input and $0.47 output per million tokens. Its model page dates the listing to August 26, 2026 and records 157 output tokens per second over the previous week.

Qwen3.8-Max is the top of the line: $2.00 input and $6.00 output per million tokens, 1M context, and no open weights. It is the tier you use when the task justifies a frontier price, not the tier you compare against a 1.7B base checkpoint.

What LFM3-1.7B-Base actually ships with

Liquid created the LFM3-1.7B-Base repository on October 26, 2026 and announced the family the following day. The checkpoint is text-only, 1.73B parameters, 24 layers of gated short convolutions and full or sliding-window attention, with a 256K-token context, a 128,000-token vocabulary, 16 named languages and roughly 1 GB of memory use at 4-bit precision. It was trained on 22 trillion tokens; the last 3.5 trillion are mid-training data that include agentic traces and tool calls and carry the context extension.

What it does not ship with is a score. The base card has no evaluation table, no context-retrieval numbers and no comparison to any other model. Every figure in Liquid's launch blog — the 49.37 average, LiveCodeBench v6 at 73.46, SWE-bench Verified at 43.40, GPQA Diamond at 63.37, IFEval at 96.66 — belongs to LFM3-1.7B after supervised fine-tuning, reinforcement learning, on-policy distillation and agentic RL. Those are real numbers and Liquid reports them honestly as its own, but they describe the finished model, not the base checkpoint whose name is in this title.

The licence is the LFM Open License v1.0. It grants commercial rights only while the licensee stays under $10 million in annual revenue; above that threshold the work is not licensed under the agreement and needs a separate deal. That single clause will remove LFM3-1.7B-Base from many shortlists before anyone runs a token through it.

Why the benchmark gap cannot be closed with arithmetic

The temptation is to line up Qwen3.8-27B's vendor numbers against Liquid's vendor numbers and declare a winner. Resist it, because the two tables do not measure the same things. Qwen reports SWE-bench Pro 61.7, Terminal Bench 2.1 at 73.0, LiveCodeBench v6 at 90.3, GPQA Diamond 89.2 and HLE 30.8. Liquid reports SWE-bench Verified 43.40, LiveCodeBench v6 73.46, GPQA Diamond 63.37 and HLE without tools at 7.17. SWE-bench Pro and SWE-bench Verified are different benchmarks with different difficulty; the parameter counts differ by a factor of 16; and one column describes a post-trained chat model while the other describes a base checkpoint that has not been post-trained at all.

There is one number worth borrowing from outside both vendors. OrcaRouter's model page for Qwen3.8-27B lists an Artificial Analysis Intelligence Index of 33.7 and an AA Coding score of 68.1, evaluated on August 14, 2026. That is an independent scoreboard, and it exists for Qwen. As of October 28, 2026, we could find no LFM3 entry on Artificial Analysis, so there is no independent counterweight on the other side. Anyone who presents a head-to-head capability verdict for these two models today is either comparing vendor tables across incompatible harnesses or inventing a number.

The deployment envelope, side by side

• Parameters — LFM3-1.7B-Base: 1.73B, text-only, 24 layers. Qwen3.8-27B: 27B dense with a vision encoder, 64 layers.

• Context — LFM3-1.7B-Base: 256K tokens. Qwen3.8-27B: 262,144 native, extensible to 1M; Qwen3.8-Flash and Qwen3.8-Max: 1M.

• Input modality — LFM3-1.7B-Base: text. Qwen3.8-27B, Qwen3.8-Flash and Qwen3.8-Max: text, image and video.

• Weights — LFM3-1.7B-Base: BF16 download with custom modelling code and trust_remote_code. Qwen3.8-27B: Apache-2.0 weights that run in Transformers, vLLM, SGLang and TokenSpeed.

• Memory floor — LFM3-1.7B-Base: about 1 GB at 4-bit, per Liquid. Qwen3.8-27B: roughly 55 GB in BF16, so a multi-GPU box or aggressive quantisation.

• Licence — LFM3-1.7B-Base: LFM Open License v1.0, commercial use below $10M annual revenue. Qwen3.8-27B: Apache 2.0, no revenue condition.

• Listed price — LFM3-1.7B-Base: no hosted per-token price published. Qwen3.8-27B on OrcaRouter: $0.33 input / $2.40 output per million tokens; Qwen3.8-Flash: $0.15 / $0.47; Qwen3.8-Max: $2.00 / $6.00.

• Observed speed — LFM3-1.7B-Base: no independent measurement. Qwen3.8-27B on OrcaRouter: 91 output tokens per second over the last seven days; Qwen3.8-Flash: 157 tokens per second.

The arithmetic that decides it

Take a workload of 100 million output tokens a month. On Qwen3.8-Flash that costs $47. On Qwen3.8-27B through OrcaRouter it costs $240. On Qwen3.8-Max it costs $600. Input tokens are cheaper on every tier, so a realistic mix lands somewhere below those figures, but output dominates the bill and the ranking does not change.

Now compare those bills with running LFM3-1.7B-Base yourself. At 4-bit it fits in about a gigabyte, which means hardware you may already own: a laptop, a recent phone, a single small GPU. The marginal cost per token is electricity and nothing else, and the model never leaves the device. Do the same arithmetic on Qwen3.8-27B and the picture flips: 100 million output tokens a month at $2.40 per million is $240, which is roughly what a rented 24 GB-class GPU costs for a month at around $0.35 an hour. Self-hosting the 27B only starts to beat the hosted price somewhere north of 100 million output tokens a month — and that calculation ignores the engineering time, the 55 GB download and the fact that a 27B model needs considerably more than 24 GB once you account for KV cache.

The practical version: below roughly 100 million output tokens a month, the hosted Qwen3.8 tiers win on total cost and operational simplicity. If your workload is small enough to fit on a phone, LFM3-1.7B-Base wins because the meter never starts. The gap between those two positions is where most real deployments live, and it is a question about volume and privacy, not about benchmark scores.

Where OrcaRouter fits, and where it does not

All three Qwen 3.8 tiers are on OrcaRouter's catalogue today, including Qwen3.8-27B self-hosted on our own infrastructure at $0.33 input and $2.40 output per million tokens, with a 262K context, a p50 time-to-first-token of 2.37 seconds and 91 output tokens per second measured over the previous week. Routing all three behind one OpenAI-compatible key is the point: a cheap default on Qwen3.8-Flash, an escalation path to Qwen3.8-27B or Qwen3.8-Max when the task needs it, and automatic failover if one upstream degrades. Because we pass provider list prices through at 0% markup, a vendor price change lands in your bill the same day rather than after a repricing cycle.

LFM3-1.7B-Base is not on our catalogue, and we are not going to pretend otherwise. If Liquid's checkpoint is the right answer for your workload, the route today is the vendor's own API and playground or a self-hosted download. What a router can still do is keep that decision reversible: put the small model behind the same interface as the hosted tiers, route by task or by cost, and move traffic back to a hosted model when the self-hosted path misbehaves.

Pick by job, not by score

Choose LFM3-1.7B-Base when the constraint is the device: offline operation, a phone or laptop, a domain-specific model you intend to post-train on a single GPU, or a workload where 256K context in a 1 GB footprint is the whole reason you are interested. Accept that you are buying raw material, that the licence has a revenue ceiling, and that no independent evaluation exists yet.

Choose Qwen3.8-27B when the job needs vision, long-horizon agents or code it can actually ship, and you have the hardware or the budget to serve 27B parameters. Choose Qwen3.8-Flash when you want most of that product surface at a fifth of the price and do not need to own the weights. Choose Qwen3.8-Max only when the task's value per token clears $6.00 of output.

The matchup in the title is a mismatch by design — a 1.7B pre-trained checkpoint against a 27B finished vision-language model — and the useful conclusion is not that one is better. It is that they answer different questions: how little hardware can I run on, versus how much capability can I buy per token. Liquid has shipped an unusually detailed answer to the first. Alibaba has been selling the second since August.