# LFM3-Omni-2B vs Qwen3.6 35B-A3B: benchmark on a DGX Spark > Liquid AI's new LFM3-Omni-2B hears, sees and speaks in about 6 GiB, and on a DGX Spark it writes 2.5 times faster than Qwen3.6 35B-A3B on the same engine. It also solves fewer than half as many coding problems and answers science questions at the level of chance. - Published: 2026-10-30 - Author: Flavien Berwick - Category: Research - Web page: https://carouter.ai/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark - In French: https://carouter.ai/fr/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark.md ![Research: LFM3-Omni-2B vs Qwen3.6 35B-A3B, benchmark on a DGX Spark. One request on llama.cpp: 160 against 65 tokens a second. Tool calls right: 64% against 81%. Science questions right: 19% against 83%. Both at 4 bits on one NVIDIA DGX Spark.](https://carouter.ai/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark/illustration.png) Liquid AI released LFM3-Omni-2B on 27 October. With 2.2 billion parameters, it takes text, images, video and audio, and it answers in text or in speech. At 4 bits it is a 1 GiB file for llama.cpp, the engine people use to run a model on their own computer. We ran it against Qwen3.6 35B-A3B, the model behind CARouter Cloud's coding endpoint, on the chip of the NVIDIA DGX Spark, which is also the chip of our own server. [The last model we tested this way: Nemotron 3.5 Bolt E4B vs Qwen3.6 35B-A3B.](https://carouter.ai/blog/nemotron-3-5-bolt-e4b-vs-qwen3-6-35b-a3b-on-a-dgx-spark) ## The short version - **Speed for one request.** LFM3-Omni-2B writes 160 tokens a second. Qwen writes 65 on the same engine, and 119 as we serve it. - **Speed for many.** With sixteen requests at once it totals 893 tokens a second, against 229 for Qwen on the same engine and 473 as we serve it. - **Memory.** With sixteen conversations open it holds 5.7 GiB, against 25.0 GiB for Qwen on the same engine. - **Quality.** It is far behind on the long tests: coding 24% against 60%, following instructions 78% against 89%, calling tools 64% against 81%, science questions 19% against 83%. - **Hearing and speaking.** Speech, images and video all worked on the few clips we tried. We did not measure how its voice sounds. ## How we compared them We did not run Qwen again. Its figures come from our earlier tests, on the same chip and with the same test programs. To check that old and new figures can sit side by side, we first ran Nemotron 3.5 Bolt E4B again. It came out 3 to 4% faster than before, so the speed ratios below lean about that far towards LFM3-Omni-2B. LFM3-Omni-2B has no reasoning mode, and Qwen reasons. The scores below compare the two models as a caller gets them. ## Speed ![One request alone on one GB10, tokens written per second. LFM3-Omni-2B on llama.cpp: 160 at 4 bits, 112 at 8 bits, 70 at 16 bits. Qwen3.6 35B-A3B: 65 at 4 bits on llama.cpp, 119 on vLLM as CARouter Cloud serves it.](https://carouter.ai/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark/01-one-request.png) For one request alone, LFM3-Omni-2B at 4 bits is 2.5 times Qwen on the same engine and 1.35 times Qwen as we serve it. Its precision matters a lot here. Going from 16 bits to 4 bits makes it 2.3 times faster, because each token then reads far less from memory. ![Sixteen requests at once on one GB10, tokens written per second in total. LFM3-Omni-2B on llama.cpp: 893 at 4 bits, 718 at 8 bits, 459 at 16 bits. Qwen3.6 35B-A3B: 229 at 4 bits on llama.cpp, 473 on vLLM as CARouter Cloud serves it.](https://carouter.ai/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark/02-sixteen-requests.png) With sixteen requests at once, the lead grows to 3.9 times Qwen on the same engine and 1.9 times Qwen as we serve it. It also reads a prompt quickly: about 13 900 tokens a second, against 2 900 for Qwen on the same engine and 6 500 as we serve it. Its tokenizer cuts our test prompt into fewer tokens, 7 069 against 11 768, so read that figure as an order of size. ## Memory and context ![Graphics memory held on one GB10. LFM3-Omni-2B on llama.cpp, image and sound parts included, with sixteen conversations open: 5.7 GiB at 4 bits, 6.4 GiB at 8 bits, 7.9 GiB at 16 bits. Qwen3.6 35B-A3B: 25.0 GiB at 4 bits on llama.cpp with sixteen conversations open, 52.6 GiB on vLLM as CARouter Cloud serves it, its whole cache reserved.](https://carouter.ai/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark/03-memory.png) On llama.cpp those figures are with sixteen conversations open. On vLLM the engine reserves its whole cache when it starts, which is what serves Qwen's longer contexts. The other side of that small size is a short context. LFM3-Omni-2B's window stops at 32 768 tokens, question and answer together, against 262 144 for Qwen. That is an eighth of the room for a codebase or a long document. ## A voice assistant We gave it the kitchen speaker from our last test: a family in Ottawa, four tools, twelve requests in a row, half of them needing a tool. It picked the right tool, or no tool, in all twelve turns. Its answer started after 0.08 seconds, against 0.21 for Qwen as we serve it with reasoning off. No listener would hear that difference, but the assistant runs in a fraction of the memory. ## Quality ![Share of each test answered correctly, both models at 4 bits, LFM3-Omni-2B without reasoning then Qwen3.6 35B-A3B with reasoning: coding 24% and 60%, following instructions 78% and 89%, calling tools 64% and 81%, calling tools without the parallel requests 75% and 81%, science questions 19% and 83%, finding a fact in a long text 100% and 100%.](https://carouter.ai/blog/lfm3-omni-2b-vs-qwen3-6-35b-a3b-on-a-dgx-spark/04-tests.png) The 4 bits are not the cause. At 16 bits LFM3-Omni-2B solved 9 coding problems and got 324 tool requests right, against 12 and 322 at 4 bits. On 50 BigCodeBench problems it solved 12 and Qwen 30. Its answers are well formatted, so the failures are real errors in the code. On 500 requests from the Berkeley function-calling benchmark it got 64% right and Qwen 81%. Part of the gap has one cause: it never made two calls in one turn, so it missed all 69 requests that need several. Qwen got 55 of those. Replaying those requests with nothing constraining its output gave the same result: one call, then it stops. In the one case we tried, it made the second call after receiving the result of the first, so an agent may get there one call per turn. Where one call is enough, it is level with Qwen: 208 against 213 of 271 requests. Where no tool fits, it calls one too often, and it was right on 114 of 160 requests against 135. IFEval gives 541 prompts with instructions a program can check, such as a word count or a required format. LFM3-Omni-2B followed every instruction of 78% of them and Qwen of 89%. On 198 graduate-level science questions (GPQA Diamond) it scored 19% and Qwen 83%. It answered 155 of them and got 37 right, which is 24%, about what guessing among four choices gives. It never finished the other 43, because it kept repeating itself until the window ran out. The same happened on 43 of the 541 IFEval prompts, with the settings Liquid recommends for text. We also hid one sentence in documents of 16 000 and 28 000 tokens and asked for it back. Both models found it 6 times out of 6. ## Hearing, seeing and speaking This is what the model is built for, and our checks here are a handful of clips rather than a benchmark. - **Speech out.** It wrote 7.5 seconds of English speech in 2.3 seconds, and 5.1 seconds of French in 2.1, loading the model included. - **Speech in.** It transcribed those two clips back with no wrong word, 30 and 24 times faster than real time. On a 10-second recording of a human reader, about 7 words in 100 were wrong. - **Images and video.** It read an invoice number and a total from an image in 0.17 seconds, and it described a short video, picture and sound, in 1.7 seconds. ## What we cannot tell you - How its voice sounds. We checked that the words came back, and no person has judged the speech. - How it holds a live spoken conversation. The connection opened, and we went no further. - How fast it is on vLLM, the engine CARouter Cloud serves with. vLLM cannot load this model yet. - Whether other settings would stop its runaway answers. We used the ones Liquid recommends. - How it handles written French. Our text tests are in English. --- More articles: [CARouter Blog](https://carouter.ai/blog.md). What CARouter is and how to call it: [llms.txt](https://carouter.ai/llms.txt).