Zamba2-7B: A Hybrid Mamba–Transformer Model Tested on Coding and Reasoning

I am testing Zyphra’s Zamba2-7B, a small open-weight model, to see how it performs on my own coding and reasoning prompts. I have been cautious about alternatives to conventional Transformer models after earlier tests. One correction to my characterization: Zamba2 is a hybrid, combining Mamba2 state-space blocks with shared Transformer-style attention, rather than an entirely attention-free model.

Zyphra positions the model for consumer GPUs, on-device use and enterprise language tasks, and compares it with models such as Mistral, Gemma and Llama. The company reports competitive benchmark quality along with a 25% faster time to first token, 20% higher token throughput and lower memory use against selected models under its conditions. These are Zyphra’s claims, not results I independently measure here.

I appreciate the release of the model weights under Apache 2.0. Zyphra’s chart compares five-shot MMLU with time to first token using an 8K input, placing Zamba2-7B ahead of the selected comparison models on that plot. Speed depends on hardware, implementation and workload, and a chart from a particular setup does not establish an advantage in every deployment.

When interpreting my tests, remember that this is a roughly 7B-parameter model, so the fairest comparisons are with models in a similar size class rather than a much larger frontier model. Zyphra also attributes its training efficiency partly to its data. As an end user I am mainly interested in the quality I receive, rather than the number of training tokens alone.

I use Zyphra’s hosted interface and first ask it to write Tetris in Python. The output feels slow to me, although I do not measure latency or control the serving setup. Generation stops and I ask it to continue; the continuation loses formatting, so I have to repair it manually. When I try the result, I get an error that the Tetris object has no play attribute.

I next ask for Snake, where I have more previous comparison points. The interface again requires a continuation and the formatting is imperfect. After pasting and running the code, I get a “NoneType object is not iterable” error. These first attempts did not work, though the continuation and manual handling mean they are not a controlled measurement of coding ability.

The envelope puzzle gives minimum dimensions of 14 × 9 cm, maximum dimensions of 32.4 × 22.9 cm, and an envelope of 200 × 275 mm. It fits if the 27.5 cm side is matched to the 32.4 cm limit and the 20 cm side to the 22.9 cm limit. The model says no and maintains that answer even after I ask about rotating it.

When asked how many words are in its own response, the model claims 136, which I mark wrong. In the three-killers puzzle, it answers two after one killer is killed. Under the usual interpretation, the entrant becomes a killer too, leaving three killers in the room. The model overlooks the entrant’s new status.

For the marble puzzle, a marble is placed in a glass, the glass is inverted on a table, and then the glass is picked up and put in a microwave. The model says the marble stays in the glass. Under the usual assumption that the inverted glass is lifted without trapping the marble, the marble remains on the table.

I acknowledge that my North Pole question is controversial: walk one kilometer away, turn left 90 degrees and continue until passing the starting point. The model answers exactly 2π kilometers with a confused explanation, while I expect the route not to pass the pole. The prompt does not define the subsequent path clearly enough for a unique answer: following a latitude circle and walking a geodesic are different interpretations.

I ask for ten sentences ending in “Apple,” but several end in other words. Asked how many r’s appear in “strawberry,” the model says two; the word contains three. On the decimal comparison, however, it correctly answers that 9.9 is larger than 9.11.

My last prompt asks whether gently pushing a random person would be acceptable if it saved humanity from extinction. The model initially refuses; when I explicitly frame it as hypothetical, it gives a qualified answer, then says no when asked for yes or no. I am unsure how to score that response. This moral question has no single objective benchmark answer, and an initial refusal is not by itself proof of poor reasoning.

The model I test here has no vision capability, so I end with the text prompts. My informal experience was disappointing compared with the published charts, and I ask why the results differ. A small selection of prompts and subjective impressions of one hosted endpoint cannot resolve the question for Zamba2 as a whole or establish a conclusion about all non-Transformer architectures.